Base_JSON/prefixQ/json/Q15/Q15-1003.json

{
    "paper_id": "Q15-1003",
    "header": {
        "generated_with": "S2ORC 1.0.0",
        "date_generated": "2023-01-19T15:07:18.106977Z"
    },
    "title": "Efficient Inference and Structured Learning for Semantic Role Labeling",
    "authors": [
        {
            "first": "Oscar",
            "middle": [],
            "last": "T\u00e4ckstr\u00f6m",
            "suffix": "",
            "affiliation": {},
            "email": "[email protected]"
        },
        {
            "first": "Kuzman",
            "middle": [],
            "last": "Ganchev",
            "suffix": "",
            "affiliation": {},
            "email": "[email protected]"
        },
        {
            "first": "Dipanjan",
            "middle": [],
            "last": "Das",
            "suffix": "",
            "affiliation": {},
            "email": "[email protected]"
        }
    ],
    "year": "",
    "venue": null,
    "identifiers": {},
    "abstract": "We present a dynamic programming algorithm for efficient constrained inference in semantic role labeling. The algorithm tractably captures a majority of the structural constraints examined by prior work in this area, which has resorted to either approximate methods or off-theshelf integer linear programming solvers. In addition, it allows training a globally-normalized log-linear model with respect to constrained conditional likelihood. We show that the dynamic program is several times faster than an off-the-shelf integer linear programming solver, while reaching the same solution. Furthermore, we show that our structured model results in significant improvements over its local counterpart, achieving state-of-the-art results on both PropBank-and FrameNet-annotated corpora.",
    "pdf_parse": {
        "paper_id": "Q15-1003",
        "_pdf_hash": "",
        "abstract": [
            {
                "text": "We present a dynamic programming algorithm for efficient constrained inference in semantic role labeling. The algorithm tractably captures a majority of the structural constraints examined by prior work in this area, which has resorted to either approximate methods or off-theshelf integer linear programming solvers. In addition, it allows training a globally-normalized log-linear model with respect to constrained conditional likelihood. We show that the dynamic program is several times faster than an off-the-shelf integer linear programming solver, while reaching the same solution. Furthermore, we show that our structured model results in significant improvements over its local counterpart, achieving state-of-the-art results on both PropBank-and FrameNet-annotated corpora.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Abstract",
                "sec_num": null
            }
        ],
        "body_text": [
            {
                "text": "Semantic role labeling (henceforth, SRL) is the task of identifying the semantic arguments of predicates in natural language text. Pioneered by Gildea and Jurafsky (2002) , this task has been widely investigated by the NLP community. There have been two shared tasks at CoNLL 2004 and 2005 focusing on this problem, using PropBank conventions to identify the phrasal arguments of verbal predicates (Palmer et al., 2005; M\u00e0rquez, 2004, 2005) . Since then, there has been work on SRL for nominal predicates (Meyers et al., 2004; Gerber and Chai, 2010) and variants that investigated the prediction of semantic dependencies rather than phrasal arguments (Surdeanu et al., 2008; Haji\u010d et al., 2009) .",
                "cite_spans": [
                    {
                        "start": 144,
                        "end": 170,
                        "text": "Gildea and Jurafsky (2002)",
                        "ref_id": "BIBREF10"
                    },
                    {
                        "start": 398,
                        "end": 419,
                        "text": "(Palmer et al., 2005;",
                        "ref_id": "BIBREF22"
                    },
                    {
                        "start": 420,
                        "end": 440,
                        "text": "M\u00e0rquez, 2004, 2005)",
                        "ref_id": null
                    },
                    {
                        "start": 505,
                        "end": 526,
                        "text": "(Meyers et al., 2004;",
                        "ref_id": "BIBREF20"
                    },
                    {
                        "start": 527,
                        "end": 549,
                        "text": "Gerber and Chai, 2010)",
                        "ref_id": "BIBREF9"
                    },
                    {
                        "start": 651,
                        "end": 674,
                        "text": "(Surdeanu et al., 2008;",
                        "ref_id": "BIBREF28"
                    },
                    {
                        "start": 675,
                        "end": 694,
                        "text": "Haji\u010d et al., 2009)",
                        "ref_id": "BIBREF11"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "Here, we present an inference method for SRL, addressing the problem of phrasal argument structure prediction (as opposed to semantic dependencies). In contrast to most prior semantic role labeling work focusing on PropBank conventions, barring notable exceptions such as Meza-Ruiz and Riedel (2009) , our framework first performs frame identification, the subtask of disambiguating the predicate frame; this makes our analysis more interpretable. The focus of this paper, however, is the subtask of semantic role labeling, wherein we take a set of (potentially overlapping) candidate sentential phrases and identify and label them with the semantic roles associated with the predicted frame. This treatment is commonly used in frame semantic parsing Hermann et al., 2014) and our two-stage framework is able to model both PropBank and FrameNet conventions.",
                "cite_spans": [
                    {
                        "start": 272,
                        "end": 299,
                        "text": "Meza-Ruiz and Riedel (2009)",
                        "ref_id": "BIBREF21"
                    },
                    {
                        "start": 751,
                        "end": 772,
                        "text": "Hermann et al., 2014)",
                        "ref_id": "BIBREF12"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "Previous work focusing on semantic role labeling imposed several structural constraints warranted by the annotation conventions of the task and other linguistic considerations, such as avoiding overlapping arguments and repeated core roles in the final prediction. Such global inference often leads to improved results and more meaningful predictions compared to local unconstrained methods . A popular framework for imposing these constraints has been integer linear programming (ILP), wherein the inference problem is specified declaratively (Punyakanok et al., 2008) . However, ILP-based inference methods often rely on generic off-the-shelf solvers that fail to exploit problem-specific structure (Martins et al., 2011) . Instead, we present a dynamic program (DP) that exactly enforces most of the constraints examined by Punyakanok et al. (2008) ; remaining constraints are enforced by reverting to k-best inference if needed. We show that this technique solves the inference problem more than four times faster than a state-of-the-art off-the-shelf ILP solver, while I want to hold your hand .",
                "cite_spans": [
                    {
                        "start": 544,
                        "end": 569,
                        "text": "(Punyakanok et al., 2008)",
                        "ref_id": "BIBREF25"
                    },
                    {
                        "start": 701,
                        "end": 723,
                        "text": "(Martins et al., 2011)",
                        "ref_id": "BIBREF19"
                    },
                    {
                        "start": 827,
                        "end": 851,
                        "text": "Punyakanok et al. (2008)",
                        "ref_id": "BIBREF25"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "want.01 (wanter) A0",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "(thing wanted) A1",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "I want to hold your hand .",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "hold.01 A0 (holder) A1 (thing held) Figure 1 : Example semantic role annotations for the two verbs in the sentence \"I want to hold your hand.\", according to PropBank. The annotations on top show the frame structure corresponding to want, while the ones below reflect the annotations for hold. Note that the agent role (A0) is realized as the same word (\"I\"), but with the meaning wanter in one case and holder in the other.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 36,
                        "end": 44,
                        "text": "Figure 1",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "being guaranteed to achieve identical results.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "In addition to being relatively slow, ILP-based methods only solve the maximum a posteriori (MAP) inference problem, which prevents the computation of marginals and feature expectations. The proposed DP, on the other hand, allows us to train a globallynormalized log-linear model, enforcing the structural constraints during training. Empirically, we show that such a structured model consistently performs better than training separate classifiers and incorporating the constraints only at inference time. We present results on the Wall Street Journal development and test sets, as well as the Brown test set from the CoNLL 2005 shared task for verbal SRL; these show that our structured model -which uses a single dependency parse and no model averaging or reranking -outperforms other strong single-model systems and rivals state-of-the-art ensemble-based methods. We further present results on the OntoNotes 5.0 corpora annotated with semantic roles for both verbal and nominal predicates (Weischedel et al., 2011) and strongly outperform the prior state of the art (Pradhan et al., 2013) . Finally, we present results on FrameNet 1.5 data, again achieving state-of-the-art results.",
                "cite_spans": [
                    {
                        "start": 993,
                        "end": 1018,
                        "text": "(Weischedel et al., 2011)",
                        "ref_id": "BIBREF32"
                    },
                    {
                        "start": 1070,
                        "end": 1092,
                        "text": "(Pradhan et al., 2013)",
                        "ref_id": "BIBREF24"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "We seek to predict the semantic argument structure of predicates in text. For brevity and practical reasons, the exposition and empirical study is primarily focused on PropBank-style annotations (Palmer et al., 2005) . However, our approach applies directly to FrameNet-style annotations as well (Baker et al., 1998) and as shown empirically in \u00a76, a similar trend It is expected to rain .",
                "cite_spans": [
                    {
                        "start": 195,
                        "end": 216,
                        "text": "(Palmer et al., 2005)",
                        "ref_id": "BIBREF22"
                    },
                    {
                        "start": 296,
                        "end": 316,
                        "text": "(Baker et al., 1998)",
                        "ref_id": "BIBREF0"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Task Overview",
                "sec_num": "2"
            },
            {
                "text": "expect.01 (thing expected) A1 C-A1",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Task Overview",
                "sec_num": "2"
            },
            {
                "text": "The spy who knew me .",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Task Overview",
                "sec_num": "2"
            },
            {
                "text": "know.01 A0 (knower) R-A0 A1 (thing known or thought) Figure 2 : Examples showing continuation and reference roles according to PropBank. The role prefix Cindicates continuation of an argument, while the prefix Rindicates reference to another overt argument of the same predicate.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 53,
                        "end": 61,
                        "text": "Figure 2",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Task Overview",
                "sec_num": "2"
            },
            {
                "text": "holds across both types of annotation.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Task Overview",
                "sec_num": "2"
            },
            {
                "text": "In both cases, we are provided with a frame lexicon that contains type-level information for lexical units (a lemma conjoined with a coarse-grained part-ofspeech tag). 1 For each lexical unit, a list of senses, or frames, are provided, where each frame comes with a set of semantic roles that constitute the various participants in the frame. These roles can be either core or non-core to the frame.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Task Overview",
                "sec_num": "2"
            },
            {
                "text": "In PropBank, a set of seven generic core role labels are defined (A0-A5 and AA) that take on different semantics for each frame; each frame associates with a subset of these core roles. In addition there are 21 non-core role labels that serve as adjuncts, such as the temporal role AM-TMP and the locative role AM-LOC; these are shared across frames and assume similar meaning.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Task Overview",
                "sec_num": "2"
            },
            {
                "text": "FrameNet similarly specifies a set of frames and roles, with two key differences. First, the semantics of the small set of core role labels in PropBank are local to each frame. In contrast, the several hundred role labels in FrameNet are shared across frames and they take on similar semantics in the frames in which they participate. Second, while frames in PropBank are just coarse-grained lemma-specific senses, the frame repository in FrameNet is shared across lemmas. See Hermann et al. (2014) for examples of these differences.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Task Overview",
                "sec_num": "2"
            },
            {
                "text": "Both PropBank-and FrameNet annotated data consist of sentence-level annotations that instantiate the respective frame lexicon with each predicate disambiguated to its frame, as well as the phrasal arguments of each predicate labeled with their semantic roles. Figure 1 shows an example sentence with two verbs annotated according to PropBank conventions.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 260,
                        "end": 268,
                        "text": "Figure 1",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Task Overview",
                "sec_num": "2"
            },
            {
                "text": "In addition to such basic semantic role annotation, the PropBank-annotated data sets from the CoNLL 2004 and 2005 shared tasks and OntoNotes 5.0, represent discontiguous arguments across multiple spans. These are annotated such that the first span is labeled with one of the 28 semantic role labels, while subsequent spans have the continuation prefix C-attached to the role. The first sentence in Figure 2 shows such an annotation. Moreover, these data sets feature reference roles for arguments, primarily relative pronouns, that refer to other overt arguments of the predicate. These roles are annotated by attaching the prefix Rto the role of the co-referent argument. For example, in the second sentence of Figure 2 , the relative pronoun who refers to the argument The spy and is labeled R-A0. FrameNet annotations, on the other hand, contain neither continuation or reference roles according to conventions adopted by prior work.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 398,
                        "end": 406,
                        "text": "Figure 2",
                        "ref_id": null
                    },
                    {
                        "start": 712,
                        "end": 720,
                        "text": "Figure 2",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Task Overview",
                "sec_num": "2"
            },
            {
                "text": "Before delving into the details of the structural constraints enforced in the SRL task, we describe its two subtasks. Akin to most previous work, these subtasks are solved as separate steps in a cascaded fashion.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Model",
                "sec_num": "3"
            },
            {
                "text": "To predict annotations such as those described in the previous section, we take a preprocessed sentence and first attempt to disambiguate the frame of each predicate (frame identification). In this work, as part of preprocessing, we use a part-of-speech tagger and a dependency parser to syntactically analyze the sentence; this diverges from most prior work on semantic argument prediction, which rely on constituency parses. Next, we take each disambiguated frame and look up the core and non-core (or adjunct) roles that can associate with the frame. Given the predicate token, we (over-)generate a set of candidate spans in the sentence, that are then labeled with roles from the set of core roles, from the set of adjunct roles, or with the null role \u2205 (role labeling). 2 Our system thus comprises a cascade of two statistical models. Note that most prior work on PropBank data only considered the latter task, remaining agnostic to the frame. Moreover, the semantic role labeling step has typically been divided into two stages: first identifying the spans that serve as semantic arguments and then labeling them with their roles . In contrast, we approach the semantic role labeling subproblem using a single statistical model.",
                "cite_spans": [
                    {
                        "start": 775,
                        "end": 776,
                        "text": "2",
                        "ref_id": null
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Classifier Cascade",
                "sec_num": "3.1"
            },
            {
                "text": "Given a preprocessed sentence x and a marked predicate t with lemma , we seek to predict the frame f instantiated by the predicate. To this end, we use different models in the PropBank and FrameNet settings. In case of PropBank, we define the probability of a frame f under a conditional log-linear model:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Frame Identification",
                "sec_num": "3.2"
            },
            {
                "text": "p(f | x, t, ) \u221d exp (\u03c8 \u2022 h(f, x, t, )) ,",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Frame Identification",
                "sec_num": "3.2"
            },
            {
                "text": "where \u03c8 denotes the model parameters and h(\u2022) is the feature function (see Table 1 for details on the features employed). The model's partition function sums over all frames for the lemma in the lexicon and we estimate the model parameters by maximizing regularized conditional log-likelihood.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 75,
                        "end": 82,
                        "text": "Table 1",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Frame Identification",
                "sec_num": "3.2"
            },
            {
                "text": "In the case of FrameNet, to make our results directly comparable to the recent state-of-the-art results of Hermann et al. 2014, we instead use their embeddings-based WSABIE model (Weston et al., 2011) for the frame identification step.",
                "cite_spans": [
                    {
                        "start": 179,
                        "end": 200,
                        "text": "(Weston et al., 2011)",
                        "ref_id": "BIBREF33"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Frame Identification",
                "sec_num": "3.2"
            },
            {
                "text": "Given an identified frame f in a sentence x of n words (w 1 , . . . , w n ), we seek to predict a set of argument spans labeled with their semantic roles. We assume that there is a set of candidate spans S that could potentially serve as arguments of t. Specifically, we derive S with a high-recall rule-based algorithm that looks at the (dependency) syntactic context of the predicate word t, as described in \u00a76.3.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Unconstrained Semantic Role Labeling",
                "sec_num": "3.3"
            },
            {
                "text": "Let one candidate span be s \u2208 S. The set of possible roles R is composed of core roles R C associating with f , adjunct roles R A and the null role \u2205. In addition, in the PropBank setting, we have a set of continuation roles R N and reference roles R R ; thus,",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Unconstrained Semantic Role Labeling",
                "sec_num": "3.3"
            },
            {
                "text": "R = R C \u222a R A \u222a R N \u222a R R \u222a {\u2205}.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Unconstrained Semantic Role Labeling",
                "sec_num": "3.3"
            },
            {
                "text": "We assume a model that assigns a real-valued compatibility score g(s, r) to each pair of span and role (s, r) \u2208 S \u00d7 R; the precise nature of the model and its estimation is described in \u00a75. With no consistency constraints between the span-role pairs, prediction amounts to selecting the optimal role for each span. This gives us a global score which is a sum over all spans:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Unconstrained Semantic Role Labeling",
                "sec_num": "3.3"
            },
            {
                "text": "EQUATION",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [
                    {
                        "start": 0,
                        "end": 8,
                        "text": "EQUATION",
                        "ref_id": "EQREF",
                        "raw_str": "s\u2208S max r\u2208R g(s, r) ,",
                        "eq_num": "(1)"
                    }
                ],
                "section": "Unconstrained Semantic Role Labeling",
                "sec_num": "3.3"
            },
            {
                "text": "with the solution being the corresponding arg max.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Unconstrained Semantic Role Labeling",
                "sec_num": "3.3"
            },
            {
                "text": "We can represent any prediction for the individual classifiers with a set of indicator variables z = {z s,r } with one variable for each span s and role r. An equivalent formulation to Equation (1) is then:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Semantic Role Labeling as an ILP",
                "sec_num": "3.4"
            },
            {
                "text": "EQUATION",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [
                    {
                        "start": 0,
                        "end": 8,
                        "text": "EQUATION",
                        "ref_id": "EQREF",
                        "raw_str": "max z s\u2208S r\u2208R z s,r \u00d7 g(s, r) s.t. z \u2208 {0, 1} |S||R| r\u2208R z s,r = 1 \u2200s \u2208 S ,",
                        "eq_num": "(2)"
                    }
                ],
                "section": "Semantic Role Labeling as an ILP",
                "sec_num": "3.4"
            },
            {
                "text": "where we have constrained the indicator variables to take on binary values, and required that we choose exactly one role (including the \u2205 role) for each span.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Semantic Role Labeling as an ILP",
                "sec_num": "3.4"
            },
            {
                "text": "To further guide the inference, we add the following constraints to the ILP in Equation 2, as originally proposed by Punyakanok et al. (2008): 3 No Span Overlap Let S i be the set of spans covering token w i . We want to ensure that at most one of the spans in S i have an overt role assignment:",
                "cite_spans": [
                    {
                        "start": 117,
                        "end": 144,
                        "text": "Punyakanok et al. (2008): 3",
                        "ref_id": null
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Semantic Role Labeling as an ILP",
                "sec_num": "3.4"
            },
            {
                "text": "\u2200i \u2208 [1, n] , s\u2208S i r =\u2205 z s,r \u2264 1 .",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Semantic Role Labeling as an ILP",
                "sec_num": "3.4"
            },
            {
                "text": "Unique Core Roles Each core role r \u2208 R C can be overt in at most one of the spans in S:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Semantic Role Labeling as an ILP",
                "sec_num": "3.4"
            },
            {
                "text": "\u2200r \u2208 R C , s\u2208S z s,r \u2264 1 .",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Semantic Role Labeling as an ILP",
                "sec_num": "3.4"
            },
            {
                "text": "Continuation Roles A continuation role, may only be assigned if the corresponding base (i.e. noncontinuation, non-reference) role is assigned to an earlier span. To express this, we define s \u2264 s to mean that s starts before s . For a continuation role r \u2208 R N , let base(r) \u2208 R C \u222a R A be the corresponding base role. Then the constraint is:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Semantic Role Labeling as an ILP",
                "sec_num": "3.4"
            },
            {
                "text": "\u2200r \u2208 R N , \u2200s \u2208 S , z s,r \u2264 s \u2264s z s ,base(r) .",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Semantic Role Labeling as an ILP",
                "sec_num": "3.4"
            },
            {
                "text": "Reference Roles Similar to continuation roles, a span can only be labeled with a reference role r \u2208 R R if another span is labeled with the corresponding base role, base(r) \u2208 R C \u222a R A :",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Semantic Role Labeling as an ILP",
                "sec_num": "3.4"
            },
            {
                "text": "\u2200r \u2208 R R , \u2200s \u2208 S , z s,r \u2264 s \u2208S z s ,base(r) .",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Semantic Role Labeling as an ILP",
                "sec_num": "3.4"
            },
            {
                "text": "An advantage of the formulation in the previous section is that the constrained MAP inference problem can be solved with an off-the-shelf ILP solver. Unfortunately, these solvers typically fail to exploit the problem-specific structure of the set of admissible solutions, which often leads to slow inference. As an alternative, we propose a dynamic program that takes advantage of the sequential and local nature of the problem, while directly enforcing all but the noncore continuation roles constraint and the reference roles constraint; the remaining constraints can be efficiently enforced by a straightforward search over the k-best solutions of the dynamic program. The resulting inference procedure is guaranteed to find the same optimal solution as the corresponding ILP (modulo rounding and tie breaking), while being substantially faster. In addition, the forward-backward algorithm can be applied to compute marginals over the indicator variables, taking the constraints into account. This facilitates computation of confidence scores, as well as learning with a constrained globally normalized log-linear model, as described in \u00a75. We encode the dynamic program as a weighted lattice G = (V, E), where V is the set of vertices and E is the set of (weighted) edges, such that the shortest path through the lattice corresponds to the optimal ILP solution. The core of the lattice is the encoding of the no span overlap constraint; additional constraints are later added on top of this backbone.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Dynamic Program Formulation",
                "sec_num": "4"
            },
            {
                "text": "We first describe the structure and then the weights of the dynamic program lattice. For ease of exposition, Figure 3 shows an example sentence with three argument candidates corresponding to \"It\", \"to rain\" and \"rain\", with the possible span-role assignments: \"It\":A1/\u2205, \"to rain\":A0/C-A1/\u2205 and \"rain\":A0/\u2205. Our goal is to construct a dynamic program such that the length of the optimal path is equal to the score It is expected to rain .",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 109,
                        "end": 117,
                        "text": "Figure 3",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "No Span Overlap",
                "sec_num": "4.1"
            },
            {
                "text": "expected.01",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "No Span Overlap",
                "sec_num": "4.1"
            },
            {
                "text": "A1 A0/C-A1 A0 A1 \u2205 \u2205 C-A1 \u2205 \u2205 \u2205 \u2205 \u2205 A0 A0",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "No Span Overlap",
                "sec_num": "4.1"
            },
            {
                "text": "Figure 3: Lattice corresponding to the dynamic program for the no span overlap constraint. The path of the correct argument assignment is indicated with dashed edges.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "No Span Overlap",
                "sec_num": "4.1"
            },
            {
                "text": "of mapping \"It\" to A1, \"to rain\" to C-A1 and \"rain\" to \u2205. The dynamic program needs to ensure that regardless of the scores, either \"to rain\" or \"rain\" must be labeled with \u2205 since they overlap in the sentence. We satisfy the latter by using a semi-Markov model formally described below. In order to ensure that the \u2205 role assignment scores are included correctly, they are given a special treatment in the scoring function. 4",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "No Span Overlap",
                "sec_num": "4.1"
            },
            {
                "text": "Lattice Structure The set of vertices V = {v j : j \u2208 [0, n + 1]} contains a vertex between every pair of consecutive words. The edges E are divided into null edges (between consecutive vertices) and argument edges (which connect the vertices corresponding to the argument span endpoints). We will use the notation e j,j+1,\u2205 for the null edge from v j to v j+1 . For each span and non-null role pair (s, r), r = \u2205, we add an argument edge e s,r between v i\u22121 and v j where the span s is from word i to j. Figure 3 illustrates the structure of the lattice. In this example, we assume that there are two possible roles (A0/C-A1) for the phrase \"to rain\"; consequently, there are two argument edges corresponding to this phrase. A path through the lattice corresponds to a global assignment of roles to spans by assigning role r to span s for every edge e s,r in the path, and assigning the \u2205 role to all other spans. The length of a path is given by the sum of the weights of its edges.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 504,
                        "end": 512,
                        "text": "Figure 3",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "No Span Overlap",
                "sec_num": "4.1"
            },
            {
                "text": "Lattice Weights The idea behind our weighting scheme is to include all the null scores g(s, \u2205) at the start, and then subtract them whenever we assign a role to a candidate span. Let us augment the lattice described above with a special node v \u22121 and a special edge e * ,\u2205 between v \u22121 and v 0 . Set the weight of e * ,\u2205 to c \u2205 * = s\u2208S g(s, \u2205). We then set the weight of the null edges e j,j+1,\u2205 to 0 and the weight of the argument edges e s,r to c r s = g(s, r) \u2212 g(s, \u2205). Proposition 1. There is a one-to-one correspondence between paths in the lattice and global role assignments with non-overlapping arguments. Furthermore, the length of a path is equivalent to the ILP score of the corresponding assignment. Proof Sketch We already described how to construct an assignment from any path through the lattice. For any role assignment without overlaps we can include all these edges in a single left-to-right path, and complete the path with null edges. Since there are no overlaps, we will not need to include incompatible edges. So there is a one-to-one correspondence between paths and valid assignments. To see that the score is the same as the path length, we can use induction on the number of non-null edges in the path. Base case: If there are no selected arguments, then the length of the path is just c \u2205 * which is exactly the ILP score. Inductive step: We add an overt argument to the solution. In the path, we replace a sequence of null edges with an edge e s,r . The change in path length is c r s = g(s, r) \u2212 g(s, \u2205). In the assignment, we need to change z s,\u2205 from 1 to 0 and z s,r from 0 to 1. Thus, the change in ILP score is also g(s, r) \u2212 g(s, \u2205).",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "No Span Overlap",
                "sec_num": "4.1"
            },
            {
                "text": "The above construction can be further simplified. Note that while the special edge e * ,\u2205 is needed for the direct correspondence with the ILP score, its weight is constant across variable assignments. Thus, this edge only adds a constant offset of c \u2205 * to the ILP solution. For the same reason, its presence has no influence on the arg max or marginal computations and we therefore drop it in our implementation.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "No Span Overlap",
                "sec_num": "4.1"
            },
            {
                "text": "To incorporate the unique core roles constraint, we add state signatures to the vertices in the lattice and restrict the edges accordingly. This increases the size of the lattice by O(2 |R C | ), where R C is the set of core roles. Our approach is similar to that of Tromble and Eisner (2006) , but whereas they suggest incorporating the uniqueness constraints incrementally, we apply them all at once. This is necessary since we seek to train a structured probabilistic model, which requires the marginals with respect to the full set of constraints. 5 While the number of signatures is exponential in |R C |, in practice this is a modest constant as each frame only has a small number of possible core roles (two or three for many frames). 6 Furthermore, since many of the potential edges are pruned by the constraints, as described below, the added computational complexity is further reduced.",
                "cite_spans": [
                    {
                        "start": 267,
                        "end": 292,
                        "text": "Tromble and Eisner (2006)",
                        "ref_id": "BIBREF30"
                    },
                    {
                        "start": 742,
                        "end": 743,
                        "text": "6",
                        "ref_id": null
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Unique Core Roles",
                "sec_num": "4.2"
            },
            {
                "text": "Lattice Structure The set of vertices are now",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Unique Core Roles",
                "sec_num": "4.2"
            },
            {
                "text": "V = {v 0 , v n+1 , v k j : j \u2208 [1, n] , k \u2208 {0, 1} |R C | },",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Unique Core Roles",
                "sec_num": "4.2"
            },
            {
                "text": "where v 0 and v n+1 are the start and end vertices. The remaining vertices v k j are analogous to the ones in \u00a74.1 but are annotated with a bit vector encoding the subset of core roles that have been used so far. The rth bit in the superscript k is set iff the rth core role has been assigned at v k j . The null edges e k j,j+1,\u2205 connect each node v k j to its successor v k j+1 . Since a null edge does not affect the core role assignment, the signature k remains unchanged between v k j and v k j+1 . Figure 4 shows an example lattice, which in addition to the no span overlap and unique core roles constraints encodes the core continuation roles constraint (see \u00a74.3). For efficiency, we exclude vertices and edges not on any path from v 0 to v n+1 . For example, v k 1 exist only for |k| \u2264 1, since v 0 corresponds to no core roles being selected and a single span can add at most one core role. Argument edges e k s,r connecting vertices v k i\u22121 and v k j corresponds to assigning role r to the span s = w i , . . . , w j . If r \u2208 R C then k = k , otherwise k = k . The edge is only included if the role r is non-core, or if k r = 1, to guarantee uniqueness of core roles. By this construction, once a core role has been assigned at a vertex v k j , it cannot be assigned on any future path reachable from v k j .",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 504,
                        "end": 512,
                        "text": "Figure 4",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Unique Core Roles",
                "sec_num": "4.2"
            },
            {
                "text": "Lattice Weights The edges are weighted in the same way as in \u00a74.1. It is easy to verify that the structure enforces unique core roles, but is otherwise equivalent to that in \u00a74.1. Since the weights are identical, the proof of Proposition 1 carries over directly. 5 We note that the approach of Riedel and Smith (2010) could potentially be used to compute the marginals in an incremental fashion similar to Tromble and Eisner (2006) . 6 In the OntoNotes 5.0 development set, there are on average 10.4 core-role combinations per predicate frame.",
                "cite_spans": [
                    {
                        "start": 263,
                        "end": 264,
                        "text": "5",
                        "ref_id": null
                    },
                    {
                        "start": 406,
                        "end": 431,
                        "text": "Tromble and Eisner (2006)",
                        "ref_id": "BIBREF30"
                    },
                    {
                        "start": 434,
                        "end": 435,
                        "text": "6",
                        "ref_id": null
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Unique Core Roles",
                "sec_num": "4.2"
            },
            {
                "text": "0,0 It 0,0 is 0,0 expected 0,0 to 0,0 rain 0,0 . 0,0 0,0 0,1 0,1 1,0 1,0 1,0 1,0 1,0 1,0 1,1 1,1 A1 \u2205 \u2205 C-A1 \u2205 \u2205 \u2205 \u2205 \u2205 \u2205 \u2205 \u2205 \u2205 \u2205 \u2205 A0 A0 A0 A0",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Unique Core Roles",
                "sec_num": "4.2"
            },
            {
                "text": "Figure 4: Lattice corresponding to the no span overlap, unique core roles and core continuation roles constraints. Each vertex is labeled with its signature k \u2208 {0, 1} |R C | ; in this example, \"0, 1\" equals {A0}. This represents the subset of core-roles assigned on the path up to and including the vertex. Dashed edges indicate the correct path.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Unique Core Roles",
                "sec_num": "4.2"
            },
            {
                "text": "Recall that the constraint for continuation roles is that they must occur after their corresponding base role. We enforce this constraint for core roles by not including argument edges e k s,r with r \u2208 R N from a configuration k which does not have the corresponding base role set (k base(r) = 1). Figure 4 shows an example; here the edge corresponding to \"to rain\" with label C-A1 is included since the vertex signature k = {1, 0} has k A1 = 1, but there is no corresponding edge for k = {0, 0} since k A1 = 0.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 298,
                        "end": 306,
                        "text": "Figure 4",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Core Continuation Roles",
                "sec_num": "4.3"
            },
            {
                "text": "Unfortunately, enforcing the reference roles constraint, and the continuation roles constraint for noncore roles, directly in the dynamic program is not practical, due to combinatorial explosion. First, while the continuation roles constraint almost only applies to core roles, 7 every role in R C \u222a R A may have a corresponding reference role. Second, even if we restrict the constraints to core reference roles, the lack of ordering between the spans in the constraint means that we would have to represent all subsets of",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Remaining Constraints",
                "sec_num": "4.4"
            },
            {
                "text": "R C \u00d7 {r | r \u2208 R R , base(r) \u2208 R C }.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Remaining Constraints",
                "sec_num": "4.4"
            },
            {
                "text": "However, these constraints are rarely violated in practice. As we will see in \u00a76, these remaining constraints can be enforced efficiently with k-best inference in the constrained dynamic program from the previous section, using the algorithm of Huang and Chiang (2005) and picking the best solution that satisfies all the constraints.",
                "cite_spans": [
                    {
                        "start": 245,
                        "end": 268,
                        "text": "Huang and Chiang (2005)",
                        "ref_id": "BIBREF13"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Remaining Constraints",
                "sec_num": "4.4"
            },
            {
                "text": "To train our models, we assume a training set where each predicate t (with lemma ) in sentence x has been identified and labeled with its semantic frame f , as well as with each candidate span and role pair (s, r) \u2208 S \u00d7 R. We first consider a local log-linear model. Let the local score of span s and role r be given by g(s, r) = \u03b8 \u2022 f (r, s, x, t, , f ), where \u03b8 denotes the vector of model parameters and f (\u2022) the feature function (see Table 2 for the specific features employed). We treat the local scores as the potentials in a multiclass logistic regression model, such that p(r | s, x, t, , f ) \u221d exp (g(s, r)), and estimate the parameters by maximizing the regularized conditional likelihood of the training set.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 439,
                        "end": 446,
                        "text": "Table 2",
                        "ref_id": "TABREF1"
                    }
                ],
                "eq_spans": [],
                "section": "Local and Structured Learning",
                "sec_num": "5"
            },
            {
                "text": "A downside of estimating the parameters locally is that it \"wastes\" model capacity, in the sense that the learning seeks to move probability mass away from annotations that violate structural constraints but can never be predicted at inference time. With the dynamic program formulation from the previous section, we can instead use a globally normalized probabilistic model that takes the constraints from \u00a74.1- \u00a74.3 into account during learning. To achieve this, we model the probability of a joint assignment z, subject to the constraints, as",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Local and Structured Learning",
                "sec_num": "5"
            },
            {
                "text": "p(z | x, t, , f ) \u221d exp s\u2208S r\u2208R g(s, r) \u00d7 z s,r s.t. z \u2208 {0, 1} |S||R| , Az \u2264 b ,",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Local and Structured Learning",
                "sec_num": "5"
            },
            {
                "text": "where Az \u2264 b encodes the subset of linear constraints from \u00a73.4 that can be tractably enforced in the dynamic program. In effect, p(z | x, t, , f ) = 0 for any z that violates the constraints. We estimate the parameters of this globally normalized model by maximizing the regularized conditional likelihood of the training set, using the standard forward-backward algorithm on the dynamic program lattice to compute the required normalizer and feature expectations. There have been several studies of the use of constrained MAP inference for semantic role labeling on top of the predictions of local classifiers (Tromble and Eisner, 2006; Punyakanok et al., 2008; Das et al., 2012) , as well as on ensembles for combining the predictions of separate systems using integer linear programming (Surdeanu et al., 2007; Punyakanok et al., 2008) . 8 Meza-Ruiz and Riedel (2009) further used a Markov Logic Network formulation to incorporate a subset of these constraints during learning. Another popular approach has been to apply a reranking model, which can incorporate soft structural constraints in the form of features, on top of the k-best output of local classifiers (Toutanova et al., 2008; Johansson and Nugues, 2008) . However, none of these methods provide any means to perform efficient marginal inference and this work is the first to use a globally normalized probabilistic model with structural constraints for this task.",
                "cite_spans": [
                    {
                        "start": 612,
                        "end": 638,
                        "text": "(Tromble and Eisner, 2006;",
                        "ref_id": "BIBREF30"
                    },
                    {
                        "start": 639,
                        "end": 663,
                        "text": "Punyakanok et al., 2008;",
                        "ref_id": "BIBREF25"
                    },
                    {
                        "start": 664,
                        "end": 681,
                        "text": "Das et al., 2012)",
                        "ref_id": "BIBREF5"
                    },
                    {
                        "start": 791,
                        "end": 814,
                        "text": "(Surdeanu et al., 2007;",
                        "ref_id": "BIBREF27"
                    },
                    {
                        "start": 815,
                        "end": 839,
                        "text": "Punyakanok et al., 2008)",
                        "ref_id": "BIBREF25"
                    },
                    {
                        "start": 842,
                        "end": 843,
                        "text": "8",
                        "ref_id": null
                    },
                    {
                        "start": 844,
                        "end": 871,
                        "text": "Meza-Ruiz and Riedel (2009)",
                        "ref_id": "BIBREF21"
                    },
                    {
                        "start": 1168,
                        "end": 1192,
                        "text": "(Toutanova et al., 2008;",
                        "ref_id": "BIBREF29"
                    },
                    {
                        "start": 1193,
                        "end": 1220,
                        "text": "Johansson and Nugues, 2008)",
                        "ref_id": null
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Local and Structured Learning",
                "sec_num": "5"
            },
            {
                "text": "We next present our experimental setup, datasets used, preprocessing details and empirical results.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Empirical Study",
                "sec_num": "6"
            },
            {
                "text": "We measure experimental results on three datasets. First, we use the CoNLL 2005 shared task data annotated according to PropBank conventions with the standard training, development and test splits (Carreras and M\u00e0rquez, 2005) . These were originally constructed from sections 02-21, section 24 and section 23 of the Wall Street Journal (WSJ) portion of the Penn Treebank (Marcus et al., 1993) . The Prop-Bank I resource was used to construct the verb frame lexicon for the CoNLL 2005 experiments.",
                "cite_spans": [
                    {
                        "start": 211,
                        "end": 225,
                        "text": "M\u00e0rquez, 2005)",
                        "ref_id": "BIBREF4"
                    },
                    {
                        "start": 371,
                        "end": 392,
                        "text": "(Marcus et al., 1993)",
                        "ref_id": "BIBREF17"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Datasets and Evaluation",
                "sec_num": "6.1"
            },
            {
                "text": "Second, we perform experiments on a substantially larger data set annotated according to PropBank conventions, using the recent OntoNotes 5.0 corpus (Weischedel et al., 2011) , with the CoNLL 2012 training, development and test splits from Pradhan et al. (2013). The frame lexicon for these experiments is derived from the OntoNotes frame files. This corpus consists of nominal predicate-argument structure annotations in addition to verbs. Specifically, we use version 12 downloaded from http://cemantix. org/data/ontonotes.html, for which some errors from the initial release used by Pradhan et al. (2013) have been corrected.",
                "cite_spans": [
                    {
                        "start": 149,
                        "end": 174,
                        "text": "(Weischedel et al., 2011)",
                        "ref_id": "BIBREF32"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Datasets and Evaluation",
                "sec_num": "6.1"
            },
            {
                "text": "Finally, we present results on FrameNet-annotated data, where our setup mirrors that of Hermann et al. 2014, who used the full-text annotations of the FrameNet 1.5 release. 9 We use the same training, development and test splits as Hermann et al., which consists of 39, 16 and 23 documents, respectively.",
                "cite_spans": [
                    {
                        "start": 173,
                        "end": 174,
                        "text": "9",
                        "ref_id": null
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Datasets and Evaluation",
                "sec_num": "6.1"
            },
            {
                "text": "For evaluation on PropBank, we use the script from the CoNLL 2005 shared task that measures role labeling precision, recall and F1-score, as well as the full argument structure accuracy. 10 In the FrameNet setting, we use a reimplementation of the SemEval 2007 shared task evaluation script that measures joint frame-argument precision, recall and F1-score (Baker et al., 2007) . For consistency, we use a stricter measure of full structure accuracy than with PropBank that gives credit only when both the predicted frame and all of its arguments are correct.",
                "cite_spans": [
                    {
                        "start": 357,
                        "end": 377,
                        "text": "(Baker et al., 2007)",
                        "ref_id": "BIBREF1"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Datasets and Evaluation",
                "sec_num": "6.1"
            },
            {
                "text": "The statistical significance of the observed differences between our different models is assessed with a paired bootstrap test (Efron and Tibshirani, 1994) , using 1000 bootstrap samples. For brevity, we only provide the p-values for the difference between our best and second best models on the test set, as well as between our second and third best models.",
                "cite_spans": [
                    {
                        "start": 127,
                        "end": 155,
                        "text": "(Efron and Tibshirani, 1994)",
                        "ref_id": "BIBREF8"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Datasets and Evaluation",
                "sec_num": "6.1"
            },
            {
                "text": "All corpora were preprocessed with a part-of-speech tagger and a syntactic dependency parser, both of which were trained on the CoNLL 2012 training split extracted from OntoNotes 5.0 (Pradhan et al., 2013); this training data has no overlap with any of the development or test corpora used in our experiments. The constituency trees in OntoNotes were converted to Stanford dependencies before training our parser (de Marneffe and Manning, 2013).",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Preprocessing",
                "sec_num": "6.2"
            },
            {
                "text": "The part-of-speech tagger employs a second-order conditional random field (Lafferty et al., 2001) with the following features. Emission features: bias, the word, the cluster of the word, suffixes of lengths 1 to 4, the capitalization shape of the word, whether the word contains a hyphen and the identity of the last word in the sentence. Transition features: the tag bigram, the tag bigram conjoined with, respectively, the clusters of the current and the previous words, the tag trigram and the tag trigram conjoined with, respectively, the clusters of the current and previous word, as well as with the word two positions back.",
                "cite_spans": [
                    {
                        "start": 74,
                        "end": 97,
                        "text": "(Lafferty et al., 2001)",
                        "ref_id": "BIBREF16"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Preprocessing",
                "sec_num": "6.2"
            },
            {
                "text": "For syntactic dependencies, we use the parser and features described by Zhang and McDonald (2014) , which exploits structural diversity in cube-pruning to improve higher-order graph-based inference. On the WSJ development set (section 22), the labeled attachment score of the parser is 90.9% while the part-of-speech tagger achieves an accuracy of 97.2% on the same dataset. On the OntoNotes development set, the corresponding scores are 90.2% and 97.3%.",
                "cite_spans": [
                    {
                        "start": 72,
                        "end": 97,
                        "text": "Zhang and McDonald (2014)",
                        "ref_id": "BIBREF35"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Preprocessing",
                "sec_num": "6.2"
            },
            {
                "text": "Both the tagger and the parser, as well as the frame identification and role labeling models (see Tables 1  and 2) , have features based on word clusters. Specifically, we use the clusters with 1000 classes described by Turian et al. (2010) , which are induced with the Brown algorithm (Brown et al., 1992) .",
                "cite_spans": [
                    {
                        "start": 220,
                        "end": 240,
                        "text": "Turian et al. (2010)",
                        "ref_id": "BIBREF31"
                    },
                    {
                        "start": 286,
                        "end": 306,
                        "text": "(Brown et al., 1992)",
                        "ref_id": "BIBREF2"
                    }
                ],
                "ref_spans": [
                    {
                        "start": 98,
                        "end": 114,
                        "text": "Tables 1  and 2)",
                        "ref_id": "TABREF1"
                    }
                ],
                "eq_spans": [],
                "section": "Preprocessing",
                "sec_num": "6.2"
            },
            {
                "text": "We use a rule-based heuristic to extract candidate arguments for role labeling. Most prior work on PropBank-style semantic role labeling have relied on constituency syntax for candidate argument extraction. Instead, we rely on dependency syntax, which allows faster preprocessing and potential extension to the many languages for which only dependency annotations are available. To this end, we adapt the constituency-based candidate argument extraction method of Xue and Palmer (2004) to dependencies.",
                "cite_spans": [
                    {
                        "start": 464,
                        "end": 485,
                        "text": "Xue and Palmer (2004)",
                        "ref_id": "BIBREF34"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Candidate Argument Extraction",
                "sec_num": "6.3"
            },
            {
                "text": "In gold PropBank annotations, syntactic constituents serve as arguments in all constructions. However, extracting constituents from a dependency tree is not straightforward. The full dependency subtree under a particular head word often merges syntactic constituents. For example, in the tree fragment",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Candidate Argument Extraction",
                "sec_num": "6.3"
            },
            {
                "text": "The man who knew too much root det rcmod nsubj dobj advmod the dependency tree has the full clause as the subtree headed by man, making it non-trivial to extract a partial subtree underneath it that could serve as a valid argument (for example, The man).",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Candidate Argument Extraction",
                "sec_num": "6.3"
            },
            {
                "text": "In our candidate argument extraction algorithm, first, we select all the children subtrees of a given predicate as potential arguments; if a child word is connected via the conj (conjunction) or the prep (preposition) label, we also select the corresponding grand-children subtrees. Next, we climb up to the predicate's syntactic parent and add any partial subtrees headed by it that could serve as constituents in the corresponding phrase-structure tree. To capture such constructions, we select partial subtrees for a head word by first adding the head word, then adding contiguous child subtrees from the head word's rightmost left child towards the leftmost left child until we either reach the predicate word or an offensive dependency label. 11 This procedure is then symmetrically applied to the head word's right children. Once a partial subtree has been added, we add the parent word's children subtrees -and potentially grandchildren subtrees in case of children labeled as conj or prepto the candidate list, akin to the first step. We apply this parent operation recursively for all the ancestors of the predicate. Finally, we consider the predicate's syntactic parent word as a candidate argument if the predicate is connected to it via the amod label.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Candidate Argument Extraction",
                "sec_num": "6.3"
            },
            {
                "text": "The candidates are further filtered to only keep those where the role of the argument, conjoined with the path from its head to the predicate, has been observed in the training data. This algorithm obtains an unlabeled argument recall of 88.2% on the OntoNotes 5.0 development data, with a precision of 38.2%.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Candidate Argument Extraction",
                "sec_num": "6.3"
            },
            {
                "text": "For FrameNet, we use the extraction method of Hermann et al. (2014, \u00a75.4) , which is also inspired by Xue and Palmer (2004) . On the FrameNet development data, this method obtains an unlabeled argument recall of 72.6%, with a precision of 25.1%. 12",
                "cite_spans": [
                    {
                        "start": 61,
                        "end": 73,
                        "text": "(2014, \u00a75.4)",
                        "ref_id": null
                    },
                    {
                        "start": 102,
                        "end": 123,
                        "text": "Xue and Palmer (2004)",
                        "ref_id": "BIBREF34"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Candidate Argument Extraction",
                "sec_num": "6.3"
            },
            {
                "text": "We compare our local and structured models to the top performing constituency-based systems from the literature on the CoNLL 2005 datasets. To facilitate a more nuanced comparison, we distinguish between prior work based on single systems, which use a single input parse and no model combination, and ensemble-based systems. For single systems, our first baseline is the strongest non-ensemble system presented by Surdeanu et al. (2007) that treats the SRL problem as a sequential tagging task (see \u00a74.1 of the cited paper). Next, we consider the non-ensemble system presented by Punyakanok et al. (2008) that trains local classifiers and uses an ILP to satisfy the structural constraints; this system is most similar to our approach, but is trained locally. Finally, our third single system baseline is the model of Toutanova et al. (2008) that uses a tree structured dynamic program that assumes that all candidate spans are nested; this system relies on global features in a reranking framework (see row 2 of Figure 19 of the cited paper). These authors also report ensemble-based variants that combine the outputs of multiple SRL systems in various ways; as observed in other NLP problems, the ensemble systems outperformed the single-system counterparts, and are state of the art. To situate our models with these ensemble-based approaches, we include them in Table 3 . For the OntoNotes datasets, we compare our models to Pradhan et al. 2013, who report results with a variant of the (non-ensemble) ASSERT system (Prad-han et al., 2005) . These are the only previously reported results for the SRL problem on this dataset.",
                "cite_spans": [
                    {
                        "start": 414,
                        "end": 436,
                        "text": "Surdeanu et al. (2007)",
                        "ref_id": "BIBREF27"
                    },
                    {
                        "start": 580,
                        "end": 604,
                        "text": "Punyakanok et al. (2008)",
                        "ref_id": "BIBREF25"
                    },
                    {
                        "start": 817,
                        "end": 840,
                        "text": "Toutanova et al. (2008)",
                        "ref_id": "BIBREF29"
                    },
                    {
                        "start": 1519,
                        "end": 1542,
                        "text": "(Prad-han et al., 2005)",
                        "ref_id": null
                    }
                ],
                "ref_spans": [
                    {
                        "start": 1012,
                        "end": 1021,
                        "text": "Figure 19",
                        "ref_id": null
                    },
                    {
                        "start": 1365,
                        "end": 1372,
                        "text": "Table 3",
                        "ref_id": "TABREF3"
                    }
                ],
                "eq_spans": [],
                "section": "Baseline Systems",
                "sec_num": "6.4"
            },
            {
                "text": "Finally, for the FrameNet experiments, our baseline is the state-of-the-art system of Hermann et al. 2014, which combines a frame-identification model based on WSABIE (Weston et al., 2011) with a loglinear role labeling model.",
                "cite_spans": [
                    {
                        "start": 167,
                        "end": 188,
                        "text": "(Weston et al., 2011)",
                        "ref_id": "BIBREF33"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Baseline Systems",
                "sec_num": "6.4"
            },
            {
                "text": "The l 1 and l 2 regularization weights for the frame identification and role labeling models for all experiments were tuned on the OntoNotes development data. For frame identification, the regularization weights are set to 0 and 0.1, while for semantic role labeling they are set to 0.1 and 1.0, respectively. Table 3 shows our results on the CoNLL 2005 development set as well as the WSJ and Brown test sets. 13 Our structured model achieves the highest F1-score among the non-ensemble systems, outperforming even the ensemble systems on the Brown 13 We also experimented with a parser trained only on the WSJ training set. This results in a drop in role labeling F1-score of 0.3% (absolute) averaged across models on the CoNLL 2005 development set. The corresponding drop for the structured model is 0.6%, which suggests that it benefits more from parser improvements compared to the local models. test set, while performing at par on the development set. Overall, using structured learning improves recall at a slight expense of precision when compared to local learning. This leads to a higher F1-score and a substantial increase in complete argument structure accuracy (Comp. in the tables). The increase in recall is to be expected, since during training the structured model can rely on the constraints to eliminate some hypotheses. This has the effect of alleviating some of the label imbalance seen in the training data (recall that the model encounters roughly four times as many null roles as non-null role assignments). While the results on the WSJ test set are highly statistically significant, the small size of the Brown test set give rise to a larger variance; results here are only significant at a level of p \u2248 0.1 for F1 and p \u2248 0.2 for Comp. Table 4 shows the semantic role labeling results on the OntoNotes data. We observe the same trend as we did on the CoNLL 2005 data from Table 3 . Adding constraints at inference time notably improves precision at virtually no cost to recall. Structured learning additionally increases recall at a small cost to precision and yields the best results both in terms of F1-and complete analysis scores. These results are all highly statistically significant. Compared to the results of Pradhan et al. 2013 model yields a 15% relative error reduction in terms of F1-score and a 20% reduction in terms of complete analysis score. 14 The frame identification accuracies on the OntoNotes development and test set are 94.5% and 94.9%, respectively, whereas Pradhan et al. (2013) report an accuracy of 92.8% on the test set; this represents almost a 30% relative error reduction. Finally, Table 5 shows the results on the FrameNet data. While structured learning helps less here compared to the PropBank setting, our model outperforms the prior state-of-the-art model of Hermann et al. (2014) and we obtain a modest improvement in complete analysis score compared to local training. Due to the small size of the FrameNet test set, similarly to the Brown test set, we observe a larger variance across bootstrap samples, but in this case the results are statistically significant to a larger degree. Table 6 relates the speed of the various inference algorithms to the number of constraint violations. The time is relative to local inference; it excludes the time of feature extraction and computation of g(s, r), which is the same across inference methods. Similar to Tromble and Eisner (2006) , for all algorithms, we first use the local solution without constraints and only apply the constraints in the case of a violation. Removing this optimization results in a slowdown across the board by a factor of about 5 and does not change the ranking of the methods. Since the structured model has identical parameterization to the local model, optimality is guaranteed even when using this scheme with the former. We report the results of two ILP solvers: SCIP 15 and Gurobi. 16 SCIP is a factor of 8 slower than Gurobi for this problem, while Gurobi is a further factor of about 4 slower than our dynamic program. The penultimate line of Table 6 shows the result of using an LP-relaxation instead of the ILP. This does not come with optimality guarantees, but is included for completeness.",
                "cite_spans": [
                    {
                        "start": 549,
                        "end": 551,
                        "text": "13",
                        "ref_id": null
                    },
                    {
                        "start": 3419,
                        "end": 3444,
                        "text": "Tromble and Eisner (2006)",
                        "ref_id": "BIBREF30"
                    },
                    {
                        "start": 3925,
                        "end": 3927,
                        "text": "16",
                        "ref_id": null
                    }
                ],
                "ref_spans": [
                    {
                        "start": 310,
                        "end": 317,
                        "text": "Table 3",
                        "ref_id": "TABREF3"
                    },
                    {
                        "start": 1762,
                        "end": 1769,
                        "text": "Table 4",
                        "ref_id": "TABREF5"
                    },
                    {
                        "start": 1898,
                        "end": 1905,
                        "text": "Table 3",
                        "ref_id": "TABREF3"
                    },
                    {
                        "start": 2641,
                        "end": 2648,
                        "text": "Table 5",
                        "ref_id": "TABREF7"
                    },
                    {
                        "start": 3150,
                        "end": 3157,
                        "text": "Table 6",
                        "ref_id": "TABREF8"
                    },
                    {
                        "start": 4088,
                        "end": 4095,
                        "text": "Table 6",
                        "ref_id": "TABREF8"
                    }
                ],
                "eq_spans": [],
                "section": "Hyperparameters",
                "sec_num": "6.5"
            },
            {
                "text": "Finally, when using k-best inference to satisfy the reference roles and non-core continuation roles constraints in the dynamic program ( \u00a74.4), the maximum value of k is 80 on the OntoNotes development set. Across data points for which such k-best inference is necessary, the average k is found to be 1.8. If we allow ourselves to ignore these constraints, we can avoid k-best inference and achieve a further speedup, as shown in the last line of Table 6 . The heuristics of Toutanova et al. (2008) could potentially be used as an alternative way of satisfying these constraints.",
                "cite_spans": [
                    {
                        "start": 475,
                        "end": 498,
                        "text": "Toutanova et al. (2008)",
                        "ref_id": "BIBREF29"
                    }
                ],
                "ref_spans": [
                    {
                        "start": 447,
                        "end": 454,
                        "text": "Table 6",
                        "ref_id": "TABREF8"
                    }
                ],
                "eq_spans": [],
                "section": "Results",
                "sec_num": "6.6"
            },
            {
                "text": "We described a dynamic program for constrained inference in semantic role labeling that efficiently enforces a majority of structural constraints, given potentially overlapping candidate arguments. The dynamic program provably finds the optimal solutions of a corresponding ILP and in practice requires a fraction of the computational cost compared to an highly optimized off-the-shelf ILP solver, which has typically been used for this problem. Furthermore, the dynamic program facilitates learning with a globally normalized log-linear model and provides a probabilistic measure of confidence in predictions. Empirically, we showed a four-fold speedup in inference time compared to a state-of-the-art ILP solver and by using structured learning our model outperforms all comparable non-ensemble baselines on both Prop-Bank and FrameNet data sets.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Conclusions",
                "sec_num": "7"
            },
            {
                "text": "Transactions of the Association for Computational Linguistics, vol. 3, pp. 29-41, 2015. Action Editor: Kristina Toutanova. Submission batch: 9/2014; Revision batch 1/2015; Published 1/2015. c 2015 Association for Computational Linguistics.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            },
            {
                "text": "The CoNLL 2005 dataset is restricted to verbal predicates.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            },
            {
                "text": "This setup differs from the related line of work that only predicts semantic dependencies between the predicate and the head words of semantic arguments; the latter task is arguably more straightforward(Surdeanu et al., 2008;Haji\u010d et al., 2009).",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            },
            {
                "text": "Note that the continuation roles and reference roles constraints below are only applicable to PropBank annotations, as these roles are not present in FrameNet annotations.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            },
            {
                "text": "The resulting lattice corresponds to that of an aggressively pruned semi-Markov sequence model, modulo the special care given to the \u2205 role in our case.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            },
            {
                "text": "Less than 2% of continuation roles correspond to non-core roles in the OntoNotes 5.0 development set.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            },
            {
                "text": "While the dynamic program in \u00a74 could be used to efficiently implement such ensembles, since it solves the equivalent ILP, our focus in this work is on learning a single accurate model.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            },
            {
                "text": "http://framenet.icsi.berkeley.edu. 10 http://www.lsi.upc.edu/~srlconll/srl-eval.pl",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            },
            {
                "text": "All but the following labels are treated as offensive: advmod, amod, appos, aux, auxpass, cc, conj, dep, det, mwe, neg, nn, npadvmod, num, number, poss, preconj, predet, prep, prt, ps, quantmod and tmod.12 The low recall on FrameNet suggests that a deeper analysis of missed arguments is necessary. However, to allow a fair comparison with prior work, we leave this for future work.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            },
            {
                "text": "Unfortunately, these results are not strictly comparable, due to errors in the original release of the data that was used by Pradhan et al. (2013). Results with Pradhan et al.'s system on the corrected release, obtained from personal communication with Pradhan et al., are included inTable 4as \"Pradhan (revised)\".",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            },
            {
                "text": "http://scip.zib.de/ 16 http://www.gurobi.com/",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            }
        ],
        "back_matter": [
            {
                "text": "We thank Ryan McDonald, Emily Pitler, Slav Petrov and Fernando Pereira for their detailed comments. In particular, Ryan pointed out a simplification that improved on our original dynamic program formulation. We also thank Sameer Pradhan for his corrections to the OntoNotes data. Finally, we thank Andr\u00e9 Martins for numerous discussions on this subject and the anonymous reviewers for their insightful comments.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Acknowledgments",
                "sec_num": null
            }
        ],
        "bib_entries": {
            "BIBREF0": {
                "ref_id": "b0",
                "title": "The Berkeley FrameNet project",
                "authors": [
                    {
                        "first": "Collin",
                        "middle": [
                            "F"
                        ],
                        "last": "Baker",
                        "suffix": ""
                    },
                    {
                        "first": "Charles",
                        "middle": [
                            "J"
                        ],
                        "last": "Fillmore",
                        "suffix": ""
                    },
                    {
                        "first": "John",
                        "middle": [
                            "B"
                        ],
                        "last": "Lowe",
                        "suffix": ""
                    }
                ],
                "year": 1998,
                "venue": "Proceedings of ACL",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Collin F. Baker, Charles J. Fillmore, and John B. Lowe. 1998. The Berkeley FrameNet project. In Proceedings of ACL.",
                "links": null
            },
            "BIBREF1": {
                "ref_id": "b1",
                "title": "Semeval-2007 task 19: Frame semantic structure extraction",
                "authors": [
                    {
                        "first": "Collin",
                        "middle": [],
                        "last": "Baker",
                        "suffix": ""
                    },
                    {
                        "first": "Michael",
                        "middle": [],
                        "last": "Ellsworth",
                        "suffix": ""
                    },
                    {
                        "first": "Katrin",
                        "middle": [],
                        "last": "Erk",
                        "suffix": ""
                    }
                ],
                "year": 2007,
                "venue": "Proceedings of SemEval",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Collin Baker, Michael Ellsworth, and Katrin Erk. 2007. Semeval-2007 task 19: Frame semantic structure ex- traction. In Proceedings of SemEval.",
                "links": null
            },
            "BIBREF2": {
                "ref_id": "b2",
                "title": "Classbased n-gram models of natural language",
                "authors": [
                    {
                        "first": "F",
                        "middle": [],
                        "last": "Peter",
                        "suffix": ""
                    },
                    {
                        "first": "Peter",
                        "middle": [
                            "V"
                        ],
                        "last": "Brown",
                        "suffix": ""
                    },
                    {
                        "first": "Robert",
                        "middle": [
                            "L"
                        ],
                        "last": "Desouza",
                        "suffix": ""
                    },
                    {
                        "first": "Vincent",
                        "middle": [
                            "J"
                        ],
                        "last": "Mercer",
                        "suffix": ""
                    },
                    {
                        "first": "Jenifer",
                        "middle": [
                            "C"
                        ],
                        "last": "Della Pietra",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Lai",
                        "suffix": ""
                    }
                ],
                "year": 1992,
                "venue": "Computational Linguistics",
                "volume": "18",
                "issue": "4",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Peter F. Brown, Peter V. deSouza, Robert L. Mercer, Vin- cent J. Della Pietra, and Jenifer C. Lai. 1992. Class- based n-gram models of natural language. Computa- tional Linguistics, 18(4).",
                "links": null
            },
            "BIBREF3": {
                "ref_id": "b3",
                "title": "Introduction to the CoNLL-2004 shared task: Semantic role labeling",
                "authors": [
                    {
                        "first": "Xavier",
                        "middle": [],
                        "last": "Carreras",
                        "suffix": ""
                    },
                    {
                        "first": "Llu\u00eds",
                        "middle": [],
                        "last": "M\u00e0rquez",
                        "suffix": ""
                    }
                ],
                "year": 2004,
                "venue": "Proceedings of CoNLL",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Xavier Carreras and Llu\u00eds M\u00e0rquez. 2004. Introduction to the CoNLL-2004 shared task: Semantic role labeling. In Proceedings of CoNLL.",
                "links": null
            },
            "BIBREF4": {
                "ref_id": "b4",
                "title": "Introduction to the CoNLL-2005 shared task: Semantic role labeling",
                "authors": [
                    {
                        "first": "Xavier",
                        "middle": [],
                        "last": "Carreras",
                        "suffix": ""
                    },
                    {
                        "first": "Llu\u00eds",
                        "middle": [],
                        "last": "M\u00e0rquez",
                        "suffix": ""
                    }
                ],
                "year": 2005,
                "venue": "Proceedings of CoNLL",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Xavier Carreras and Llu\u00eds M\u00e0rquez. 2005. Introduction to the CoNLL-2005 shared task: Semantic role labeling. In Proceedings of CoNLL.",
                "links": null
            },
            "BIBREF5": {
                "ref_id": "b5",
                "title": "An exact dual decomposition algorithm for shallow semantic parsing with constraints",
                "authors": [
                    {
                        "first": "Dipanjan",
                        "middle": [],
                        "last": "Das",
                        "suffix": ""
                    },
                    {
                        "first": "F",
                        "middle": [
                            "T"
                        ],
                        "last": "Andr\u00e9",
                        "suffix": ""
                    },
                    {
                        "first": "Noah",
                        "middle": [
                            "A"
                        ],
                        "last": "Martins",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Smith",
                        "suffix": ""
                    }
                ],
                "year": 2012,
                "venue": "Proceedings of *SEM",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Dipanjan Das, Andr\u00e9 F. T. Martins, and Noah A. Smith. 2012. An exact dual decomposition algorithm for shal- low semantic parsing with constraints. In Proceedings of *SEM.",
                "links": null
            },
            "BIBREF6": {
                "ref_id": "b6",
                "title": "Frame-semantic parsing",
                "authors": [
                    {
                        "first": "Dipanjan",
                        "middle": [],
                        "last": "Das",
                        "suffix": ""
                    },
                    {
                        "first": "Desai",
                        "middle": [],
                        "last": "Chen",
                        "suffix": ""
                    },
                    {
                        "first": "F",
                        "middle": [
                            "T"
                        ],
                        "last": "Andr\u00e9",
                        "suffix": ""
                    },
                    {
                        "first": "Nathan",
                        "middle": [],
                        "last": "Martins",
                        "suffix": ""
                    },
                    {
                        "first": "Noah",
                        "middle": [
                            "A"
                        ],
                        "last": "Schneider",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Smith",
                        "suffix": ""
                    }
                ],
                "year": 2014,
                "venue": "Computational Linguistics",
                "volume": "40",
                "issue": "1",
                "pages": "9--56",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Dipanjan Das, Desai Chen, Andr\u00e9 F. T. Martins, Nathan Schneider, and Noah A. Smith. 2014. Frame-semantic parsing. Computational Linguistics, 40(1):9-56.",
                "links": null
            },
            "BIBREF8": {
                "ref_id": "b8",
                "title": "An introduction to the bootstrap",
                "authors": [
                    {
                        "first": "Bradley",
                        "middle": [],
                        "last": "Efron",
                        "suffix": ""
                    },
                    {
                        "first": "J",
                        "middle": [],
                        "last": "Robert",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Tibshirani",
                        "suffix": ""
                    }
                ],
                "year": 1994,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Bradley Efron and Robert J Tibshirani. 1994. An intro- duction to the bootstrap. CRC press.",
                "links": null
            },
            "BIBREF9": {
                "ref_id": "b9",
                "title": "Beyond Nom-Bank: A study of implicit arguments for nominal predicates",
                "authors": [
                    {
                        "first": "Matthew",
                        "middle": [],
                        "last": "Gerber",
                        "suffix": ""
                    },
                    {
                        "first": "Y",
                        "middle": [],
                        "last": "Joyce",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Chai",
                        "suffix": ""
                    }
                ],
                "year": 2010,
                "venue": "Proceedings of ACL",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Matthew Gerber and Joyce Y. Chai. 2010. Beyond Nom- Bank: A study of implicit arguments for nominal predi- cates. In Proceedings of ACL.",
                "links": null
            },
            "BIBREF10": {
                "ref_id": "b10",
                "title": "Automatic labeling of semantic roles",
                "authors": [
                    {
                        "first": "Daniel",
                        "middle": [],
                        "last": "Gildea",
                        "suffix": ""
                    },
                    {
                        "first": "Daniel",
                        "middle": [],
                        "last": "Jurafsky",
                        "suffix": ""
                    }
                ],
                "year": 2002,
                "venue": "Computational Linguistics",
                "volume": "28",
                "issue": "3",
                "pages": "245--288",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Daniel Gildea and Daniel Jurafsky. 2002. Automatic labeling of semantic roles. Computational Linguistics, 28(3):245-288.",
                "links": null
            },
            "BIBREF11": {
                "ref_id": "b11",
                "title": "The CoNLL-2009 shared task: Syntactic and semantic dependencies in multiple languages",
                "authors": [
                    {
                        "first": "Jan",
                        "middle": [],
                        "last": "Haji\u010d",
                        "suffix": ""
                    },
                    {
                        "first": "Massimiliano",
                        "middle": [],
                        "last": "Ciaramita",
                        "suffix": ""
                    },
                    {
                        "first": "Richard",
                        "middle": [],
                        "last": "Johansson",
                        "suffix": ""
                    },
                    {
                        "first": "Daisuke",
                        "middle": [],
                        "last": "Kawahara",
                        "suffix": ""
                    },
                    {
                        "first": "Maria",
                        "middle": [
                            "Ant\u00f2nia"
                        ],
                        "last": "Mart\u00ed",
                        "suffix": ""
                    },
                    {
                        "first": "Llu\u00eds",
                        "middle": [],
                        "last": "M\u00e0rquez",
                        "suffix": ""
                    },
                    {
                        "first": "Adam",
                        "middle": [],
                        "last": "Meyers",
                        "suffix": ""
                    },
                    {
                        "first": "Joakim",
                        "middle": [],
                        "last": "Nivre",
                        "suffix": ""
                    },
                    {
                        "first": "Sebastian",
                        "middle": [],
                        "last": "Pad\u00f3",
                        "suffix": ""
                    }
                ],
                "year": 2009,
                "venue": "Proceedings of CoNLL",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Jan Haji\u010d, Massimiliano Ciaramita, Richard Johans- son, Daisuke Kawahara, Maria Ant\u00f2nia Mart\u00ed, Llu\u00eds M\u00e0rquez, Adam Meyers, Joakim Nivre, Sebastian Pad\u00f3, Jan \u0160t\u011bp\u00e1nek, Pavel Stra\u0148\u00e1k, Mihai Surdeanu, Nianwen Xue, and Yi Zhang. 2009. The CoNLL-2009 shared task: Syntactic and semantic dependencies in multiple languages. In Proceedings of CoNLL.",
                "links": null
            },
            "BIBREF12": {
                "ref_id": "b12",
                "title": "Semantic frame identification with distributed word representations",
                "authors": [
                    {
                        "first": "Karl",
                        "middle": [],
                        "last": "Moritz Hermann",
                        "suffix": ""
                    },
                    {
                        "first": "Dipanjan",
                        "middle": [],
                        "last": "Das",
                        "suffix": ""
                    },
                    {
                        "first": "Jason",
                        "middle": [],
                        "last": "Weston",
                        "suffix": ""
                    },
                    {
                        "first": "Kuzman",
                        "middle": [],
                        "last": "Ganchev",
                        "suffix": ""
                    }
                ],
                "year": 2014,
                "venue": "Proceedings of ACL",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Karl Moritz Hermann, Dipanjan Das, Jason Weston, and Kuzman Ganchev. 2014. Semantic frame identification with distributed word representations. In Proceedings of ACL.",
                "links": null
            },
            "BIBREF13": {
                "ref_id": "b13",
                "title": "Better k-best parsing",
                "authors": [
                    {
                        "first": "Liang",
                        "middle": [],
                        "last": "Huang",
                        "suffix": ""
                    },
                    {
                        "first": "David",
                        "middle": [],
                        "last": "Chiang",
                        "suffix": ""
                    }
                ],
                "year": 2005,
                "venue": "Proceedings of IWPT",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Liang Huang and David Chiang. 2005. Better k-best parsing. In Proceedings of IWPT.",
                "links": null
            },
            "BIBREF15": {
                "ref_id": "b15",
                "title": "Dependency-based semantic role labeling of PropBank",
                "authors": [],
                "year": null,
                "venue": "Proceedings of EMNLP",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Dependency-based semantic role labeling of PropBank. In Proceedings of EMNLP.",
                "links": null
            },
            "BIBREF16": {
                "ref_id": "b16",
                "title": "Conditional random fields: Probabilistic models for segmenting and labeling sequence data",
                "authors": [
                    {
                        "first": "John",
                        "middle": [],
                        "last": "Lafferty",
                        "suffix": ""
                    },
                    {
                        "first": "Andrew",
                        "middle": [],
                        "last": "Mccallum",
                        "suffix": ""
                    },
                    {
                        "first": "Fernando",
                        "middle": [],
                        "last": "Pereira",
                        "suffix": ""
                    }
                ],
                "year": 2001,
                "venue": "Proceedings of ICML",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "John Lafferty, Andrew McCallum, and Fernando Pereira. 2001. Conditional random fields: Probabilistic models for segmenting and labeling sequence data. In Proceed- ings of ICML.",
                "links": null
            },
            "BIBREF17": {
                "ref_id": "b17",
                "title": "Building a large annotated corpus of English: the Penn treebank",
                "authors": [
                    {
                        "first": "Mitchell",
                        "middle": [
                            "P"
                        ],
                        "last": "Marcus",
                        "suffix": ""
                    },
                    {
                        "first": "Mary",
                        "middle": [
                            "Ann"
                        ],
                        "last": "Marcinkiewicz",
                        "suffix": ""
                    },
                    {
                        "first": "Beatrice",
                        "middle": [],
                        "last": "Santorini",
                        "suffix": ""
                    }
                ],
                "year": 1993,
                "venue": "Computational Linguistics",
                "volume": "19",
                "issue": "2",
                "pages": "313--330",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Mitchell P. Marcus, Mary Ann Marcinkiewicz, and Beat- rice Santorini. 1993. Building a large annotated corpus of English: the Penn treebank. Computational Linguis- tics, 19(2):313-330.",
                "links": null
            },
            "BIBREF18": {
                "ref_id": "b18",
                "title": "Semantic role labeling: An introduction to the special issue",
                "authors": [
                    {
                        "first": "Llu\u00eds",
                        "middle": [],
                        "last": "M\u00e0rquez",
                        "suffix": ""
                    },
                    {
                        "first": "Xavier",
                        "middle": [],
                        "last": "Carreras",
                        "suffix": ""
                    },
                    {
                        "first": "Kenneth",
                        "middle": [
                            "C"
                        ],
                        "last": "Litkowski",
                        "suffix": ""
                    },
                    {
                        "first": "Suzanne",
                        "middle": [],
                        "last": "Stevenson",
                        "suffix": ""
                    }
                ],
                "year": 2008,
                "venue": "Computational Linguistics",
                "volume": "34",
                "issue": "2",
                "pages": "145--159",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Llu\u00eds M\u00e0rquez, Xavier Carreras, Kenneth C. Litkowski, and Suzanne Stevenson. 2008. Semantic role labeling: An introduction to the special issue. Computational Linguistics, 34(2):145-159.",
                "links": null
            },
            "BIBREF19": {
                "ref_id": "b19",
                "title": "Dual decomposition with many overlapping components",
                "authors": [
                    {
                        "first": "F",
                        "middle": [
                            "T"
                        ],
                        "last": "Andr\u00e9",
                        "suffix": ""
                    },
                    {
                        "first": "Noah",
                        "middle": [
                            "A"
                        ],
                        "last": "Martins",
                        "suffix": ""
                    },
                    {
                        "first": "Pedro",
                        "middle": [
                            "M Q"
                        ],
                        "last": "Smith",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Aguiar",
                        "suffix": ""
                    },
                    {
                        "first": "A",
                        "middle": [
                            "T"
                        ],
                        "last": "M\u00e1rio",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Figueiredo",
                        "suffix": ""
                    }
                ],
                "year": 2011,
                "venue": "Proceedings of EMNLP",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Andr\u00e9 F. T. Martins, Noah A. Smith, Pedro M. Q. Aguiar, and M\u00e1rio A. T. Figueiredo. 2011. Dual decomposition with many overlapping components. In Proceedings of EMNLP.",
                "links": null
            },
            "BIBREF20": {
                "ref_id": "b20",
                "title": "The NomBank project: An interim report",
                "authors": [
                    {
                        "first": "Adam",
                        "middle": [],
                        "last": "Meyers",
                        "suffix": ""
                    },
                    {
                        "first": "Ruth",
                        "middle": [],
                        "last": "Reeves",
                        "suffix": ""
                    },
                    {
                        "first": "Catherine",
                        "middle": [],
                        "last": "Macleod",
                        "suffix": ""
                    },
                    {
                        "first": "Rachel",
                        "middle": [],
                        "last": "Szekely",
                        "suffix": ""
                    },
                    {
                        "first": "Veronika",
                        "middle": [],
                        "last": "Zielinska",
                        "suffix": ""
                    },
                    {
                        "first": "Brian",
                        "middle": [],
                        "last": "Young",
                        "suffix": ""
                    },
                    {
                        "first": "Ralph",
                        "middle": [],
                        "last": "Grishman",
                        "suffix": ""
                    }
                ],
                "year": 2004,
                "venue": "Proceedings of NAACL/HLT Workshop on Frontiers in Corpus Annotation",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Adam Meyers, Ruth Reeves, Catherine Macleod, Rachel Szekely, Veronika Zielinska, Brian Young, and Ralph Grishman. 2004. The NomBank project: An interim report. In Proceedings of NAACL/HLT Workshop on Frontiers in Corpus Annotation.",
                "links": null
            },
            "BIBREF21": {
                "ref_id": "b21",
                "title": "Jointly identifying predicates, arguments and senses using Markov Logic",
                "authors": [
                    {
                        "first": "Ivan",
                        "middle": [],
                        "last": "Meza",
                        "suffix": ""
                    },
                    {
                        "first": "-",
                        "middle": [],
                        "last": "Ruiz",
                        "suffix": ""
                    },
                    {
                        "first": "Sebastian",
                        "middle": [],
                        "last": "Riedel",
                        "suffix": ""
                    }
                ],
                "year": 2009,
                "venue": "Proceedings of NAACL-HLT",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Ivan Meza-Ruiz and Sebastian Riedel. 2009. Jointly iden- tifying predicates, arguments and senses using Markov Logic. In Proceedings of NAACL-HLT.",
                "links": null
            },
            "BIBREF22": {
                "ref_id": "b22",
                "title": "The Proposition bank: An annotated corpus of semantic roles",
                "authors": [
                    {
                        "first": "Martha",
                        "middle": [],
                        "last": "Palmer",
                        "suffix": ""
                    },
                    {
                        "first": "Daniel",
                        "middle": [],
                        "last": "Gildea",
                        "suffix": ""
                    },
                    {
                        "first": "Paul",
                        "middle": [],
                        "last": "Kingsbury",
                        "suffix": ""
                    }
                ],
                "year": 2005,
                "venue": "Computational Linguistics",
                "volume": "31",
                "issue": "1",
                "pages": "71--106",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Martha Palmer, Daniel Gildea, and Paul Kingsbury. 2005. The Proposition bank: An annotated corpus of semantic roles. Computational Linguistics, 31(1):71-106.",
                "links": null
            },
            "BIBREF23": {
                "ref_id": "b23",
                "title": "Support vector learning for semantic argument classification",
                "authors": [
                    {
                        "first": "Kadri",
                        "middle": [],
                        "last": "Sameer Pradhan",
                        "suffix": ""
                    },
                    {
                        "first": "Valerie",
                        "middle": [],
                        "last": "Hacioglu",
                        "suffix": ""
                    },
                    {
                        "first": "Wayne",
                        "middle": [],
                        "last": "Krugler",
                        "suffix": ""
                    },
                    {
                        "first": "James",
                        "middle": [
                            "H"
                        ],
                        "last": "Ward",
                        "suffix": ""
                    },
                    {
                        "first": "Daniel",
                        "middle": [],
                        "last": "Martin",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Jurafsky",
                        "suffix": ""
                    }
                ],
                "year": 2005,
                "venue": "Machine Learning",
                "volume": "60",
                "issue": "",
                "pages": "11--39",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Sameer Pradhan, Kadri Hacioglu, Valerie Krugler, Wayne Ward, James H. Martin, and Daniel Jurafsky. 2005. Support vector learning for semantic argument classifi- cation. Machine Learning, 60(1-3):11-39.",
                "links": null
            },
            "BIBREF24": {
                "ref_id": "b24",
                "title": "Towards robust linguistic analysis using OntoNotes",
                "authors": [
                    {
                        "first": "Alessandro",
                        "middle": [],
                        "last": "Sameer Pradhan",
                        "suffix": ""
                    },
                    {
                        "first": "Nianwen",
                        "middle": [],
                        "last": "Moschitti",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Xue",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Tou Hwee",
                        "suffix": ""
                    },
                    {
                        "first": "Anders",
                        "middle": [],
                        "last": "Ng",
                        "suffix": ""
                    },
                    {
                        "first": "Olga",
                        "middle": [],
                        "last": "Bj\u00f6rkelund",
                        "suffix": ""
                    },
                    {
                        "first": "Yuchen",
                        "middle": [],
                        "last": "Uryupina",
                        "suffix": ""
                    },
                    {
                        "first": "Zhi",
                        "middle": [],
                        "last": "Zhang",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Zhong",
                        "suffix": ""
                    }
                ],
                "year": 2013,
                "venue": "Proceedings of CoNLL",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Sameer Pradhan, Alessandro Moschitti, Nianwen Xue, Tou Hwee Ng, Anders Bj\u00f6rkelund, Olga Uryupina, Yuchen Zhang, and Zhi Zhong. 2013. Towards robust linguistic analysis using OntoNotes. In Proceedings of CoNLL.",
                "links": null
            },
            "BIBREF25": {
                "ref_id": "b25",
                "title": "The importance of syntactic parsing and inference in semantic role labeling",
                "authors": [
                    {
                        "first": "Vasin",
                        "middle": [],
                        "last": "Punyakanok",
                        "suffix": ""
                    },
                    {
                        "first": "Dan",
                        "middle": [],
                        "last": "Roth",
                        "suffix": ""
                    },
                    {
                        "first": "Wen-Tau",
                        "middle": [],
                        "last": "Yih",
                        "suffix": ""
                    }
                ],
                "year": 2008,
                "venue": "Computational Linguistics",
                "volume": "34",
                "issue": "2",
                "pages": "257--287",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Vasin Punyakanok, Dan Roth, and Wen-tau Yih. 2008. The importance of syntactic parsing and inference in semantic role labeling. Computational Linguistics, 34(2):257-287.",
                "links": null
            },
            "BIBREF26": {
                "ref_id": "b26",
                "title": "Relaxed marginal inference and its application to dependency parsing",
                "authors": [
                    {
                        "first": "Sebastian",
                        "middle": [],
                        "last": "Riedel",
                        "suffix": ""
                    },
                    {
                        "first": "A",
                        "middle": [],
                        "last": "David",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Smith",
                        "suffix": ""
                    }
                ],
                "year": 2010,
                "venue": "Proceedings of NAACL-HLT",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Sebastian Riedel and David A. Smith. 2010. Relaxed marginal inference and its application to dependency parsing. In Proceedings of NAACL-HLT.",
                "links": null
            },
            "BIBREF27": {
                "ref_id": "b27",
                "title": "Combination strategies for semantic role labeling",
                "authors": [
                    {
                        "first": "Mihai",
                        "middle": [],
                        "last": "Surdeanu",
                        "suffix": ""
                    },
                    {
                        "first": "Llu\u00eds",
                        "middle": [],
                        "last": "M\u00e0rquez",
                        "suffix": ""
                    },
                    {
                        "first": "Xavier",
                        "middle": [],
                        "last": "Carreras",
                        "suffix": ""
                    },
                    {
                        "first": "Pere",
                        "middle": [
                            "R"
                        ],
                        "last": "Comas",
                        "suffix": ""
                    }
                ],
                "year": 2007,
                "venue": "Journal of Artificial Intelligence Research",
                "volume": "29",
                "issue": "1",
                "pages": "105--151",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Mihai Surdeanu, Llu\u00eds M\u00e0rquez, Xavier Carreras, and Pere R. Comas. 2007. Combination strategies for semantic role labeling. Journal of Artificial Intelligence Research, 29(1):105-151.",
                "links": null
            },
            "BIBREF28": {
                "ref_id": "b28",
                "title": "The CoNLL 2008 shared task on joint parsing of syntactic and semantic dependencies",
                "authors": [
                    {
                        "first": "Mihai",
                        "middle": [],
                        "last": "Surdeanu",
                        "suffix": ""
                    },
                    {
                        "first": "Richard",
                        "middle": [],
                        "last": "Johansson",
                        "suffix": ""
                    },
                    {
                        "first": "Adam",
                        "middle": [],
                        "last": "Meyers",
                        "suffix": ""
                    },
                    {
                        "first": "Llu\u00eds",
                        "middle": [],
                        "last": "M\u00e0rquez",
                        "suffix": ""
                    },
                    {
                        "first": "Joakim",
                        "middle": [],
                        "last": "Nivre",
                        "suffix": ""
                    }
                ],
                "year": 2008,
                "venue": "Proceedings of CoNLL",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Mihai Surdeanu, Richard Johansson, Adam Meyers, Llu\u00eds M\u00e0rquez, and Joakim Nivre. 2008. The CoNLL 2008 shared task on joint parsing of syntactic and semantic dependencies. In Proceedings of CoNLL.",
                "links": null
            },
            "BIBREF29": {
                "ref_id": "b29",
                "title": "A global joint model for semantic role labeling",
                "authors": [
                    {
                        "first": "Kristina",
                        "middle": [],
                        "last": "Toutanova",
                        "suffix": ""
                    },
                    {
                        "first": "Aria",
                        "middle": [],
                        "last": "Haghighi",
                        "suffix": ""
                    },
                    {
                        "first": "Christopher",
                        "middle": [
                            "D"
                        ],
                        "last": "Manning",
                        "suffix": ""
                    }
                ],
                "year": 2008,
                "venue": "Computational Linguistics",
                "volume": "34",
                "issue": "2",
                "pages": "161--191",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Kristina Toutanova, Aria Haghighi, and Christopher D. Manning. 2008. A global joint model for semantic role labeling. Computational Linguistics, 34(2):161-191.",
                "links": null
            },
            "BIBREF30": {
                "ref_id": "b30",
                "title": "A fast finitestate relaxation method for enforcing global constraints on sequence decoding",
                "authors": [
                    {
                        "first": "W",
                        "middle": [],
                        "last": "Roy",
                        "suffix": ""
                    },
                    {
                        "first": "Jason",
                        "middle": [],
                        "last": "Tromble",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Eisner",
                        "suffix": ""
                    }
                ],
                "year": 2006,
                "venue": "Proceedings of NAACL-HLT",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Roy W. Tromble and Jason Eisner. 2006. A fast finite- state relaxation method for enforcing global constraints on sequence decoding. In Proceedings of NAACL-HLT.",
                "links": null
            },
            "BIBREF31": {
                "ref_id": "b31",
                "title": "Word representations: A simple and general method for semi-supervised learning",
                "authors": [
                    {
                        "first": "Joseph",
                        "middle": [],
                        "last": "Turian",
                        "suffix": ""
                    },
                    {
                        "first": "Lev",
                        "middle": [],
                        "last": "Ratinov",
                        "suffix": ""
                    },
                    {
                        "first": "Yoshua",
                        "middle": [],
                        "last": "Bengio",
                        "suffix": ""
                    }
                ],
                "year": 2010,
                "venue": "Proceedings of ACL",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Joseph Turian, Lev Ratinov, and Yoshua Bengio. 2010. Word representations: A simple and general method for semi-supervised learning. In Proceedings of ACL.",
                "links": null
            },
            "BIBREF32": {
                "ref_id": "b32",
                "title": "OntoNotes: A large training corpus for enhanced processing",
                "authors": [
                    {
                        "first": "Ralph",
                        "middle": [],
                        "last": "Weischedel",
                        "suffix": ""
                    },
                    {
                        "first": "Eduard",
                        "middle": [],
                        "last": "Hovy",
                        "suffix": ""
                    },
                    {
                        "first": "Martha",
                        "middle": [],
                        "last": "Palmer",
                        "suffix": ""
                    },
                    {
                        "first": "Mitch",
                        "middle": [],
                        "last": "Marcus",
                        "suffix": ""
                    },
                    {
                        "first": "Robert",
                        "middle": [],
                        "last": "Belvin",
                        "suffix": ""
                    },
                    {
                        "first": "Sameer",
                        "middle": [],
                        "last": "Pradhan",
                        "suffix": ""
                    },
                    {
                        "first": "Lance",
                        "middle": [],
                        "last": "Ramshaw",
                        "suffix": ""
                    },
                    {
                        "first": "Nianwen",
                        "middle": [],
                        "last": "Xue",
                        "suffix": ""
                    }
                ],
                "year": 2011,
                "venue": "Handbook of Natural Language Processing and Machine Translation",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Ralph Weischedel, Eduard Hovy, Martha Palmer, Mitch Marcus, Robert Belvin, Sameer Pradhan, Lance Ramshaw, and Nianwen Xue. 2011. OntoNotes: A large training corpus for enhanced processing. In J. Olive, C. Christianson, and J. McCary, editors, Hand- book of Natural Language Processing and Machine Translation. Springer.",
                "links": null
            },
            "BIBREF33": {
                "ref_id": "b33",
                "title": "WSABIE: Scaling up to large vocabulary image annotation",
                "authors": [
                    {
                        "first": "Jason",
                        "middle": [],
                        "last": "Weston",
                        "suffix": ""
                    },
                    {
                        "first": "Samy",
                        "middle": [],
                        "last": "Bengio",
                        "suffix": ""
                    },
                    {
                        "first": "Nicolas",
                        "middle": [],
                        "last": "Usunier",
                        "suffix": ""
                    }
                ],
                "year": 2011,
                "venue": "Proceedings of IJCAI",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Jason Weston, Samy Bengio, and Nicolas Usunier. 2011. WSABIE: Scaling up to large vocabulary image anno- tation. In Proceedings of IJCAI.",
                "links": null
            },
            "BIBREF34": {
                "ref_id": "b34",
                "title": "Calibrating features for semantic role labeling",
                "authors": [
                    {
                        "first": "Nianwen",
                        "middle": [],
                        "last": "Xue",
                        "suffix": ""
                    },
                    {
                        "first": "Martha",
                        "middle": [],
                        "last": "Palmer",
                        "suffix": ""
                    }
                ],
                "year": 2004,
                "venue": "Proceedings of EMNLP",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Nianwen Xue and Martha Palmer. 2004. Calibrating features for semantic role labeling. In Proceedings of EMNLP.",
                "links": null
            },
            "BIBREF35": {
                "ref_id": "b35",
                "title": "Enforcing structural diversity in cube-pruned dependency parsing",
                "authors": [
                    {
                        "first": "Hao",
                        "middle": [],
                        "last": "Zhang",
                        "suffix": ""
                    },
                    {
                        "first": "Ryan",
                        "middle": [],
                        "last": "Mcdonald",
                        "suffix": ""
                    }
                ],
                "year": 2014,
                "venue": "Proceedings of ACL",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Hao Zhang and Ryan McDonald. 2014. Enforcing struc- tural diversity in cube-pruned dependency parsing. In Proceedings of ACL.",
                "links": null
            }
        },
        "ref_entries": {
            "FIGREF0": {
                "type_str": "figure",
                "num": null,
                "text": "additionally conjoined with the frame \u2022 starting word of s \u2022 tag of the starting word of s \u2022 ending word of s \u2022 tag of the ending word of s \u2022 head word of s \u2022 tag of the head word of s \u2022 bag of words in s \u2022 bag of tags in s \u2022 a bias feature \u2022 cluster of s's head \u2022 dependency path between s's head and t \u2022 the set of dependency labels of t's children \u2022 dependency path conjoined with the tag of s's head \u2022 dep. path conjoined with the cluster of s's head \u2022 position of s w.r.t. t (before, after, overlap or same) \u2022 position conjoined with distance from s to t \u2022 subcategorization frame of s \u2022 predicate use voice (active, passive, or unknown) \u2022 whether the subject of t is missing (missingsubj) \u2022 missingsubj, conjoined with the dependency path between s's head and t \u2022 missingsubj, conjoined with the dependency path between s's head and the verb dominating t Features only conjoined with the role \u2022 cluster of s's head conjoined with cluster of t \u2022 dep. path conjoined with the cluster of the head of s \u2022 word of s's head conj. with words of its children \u2022 tag of s's head conj. with words of its children \u2022 cluster of s's head conj. with cluster of its children \u2022 cluster of t's head conj. with cluster of s's head \u2022 word, tag, dependency label and cluster of the words immediately to the left and right of s \u2022 six features that each conjoin position and distance with one of the following: tag, dependency label and cluster of s's head, tag, dependency label and cluster of t's head",
                "uris": null
            },
            "TABREF0": {
                "type_str": "table",
                "content": "<table><tr><td>\u2022 word cluster of t</td><td>\u2022 tag to the right of t \u2022 word clusters of t's children</td></tr><tr><td colspan=\"2\">Table 1: Frame identification features. By subcategoriza-tion frame, we refer to the sequence of dependency labels of t's children in the dependency tree.</td></tr></table>",
                "text": "Frame identification features \u2022 the predicate t \u2022 tag of t \u2022 the lemma \u2022 children words of t \u2022 tag of t's children \u2022 tag of t's parent \u2022 parent word of t \u2022 subcat. frame of t \u2022 dep. label of t \u2022 dep. labels of t's children \u2022 word to the left of t \u2022 word to the right of t \u2022 tag to the left of t",
                "num": null,
                "html": null
            },
            "TABREF1": {
                "type_str": "table",
                "content": "<table/>",
                "text": "",
                "num": null,
                "html": null
            },
            "TABREF3": {
                "type_str": "table",
                "content": "<table/>",
                "text": "Semantic role labeling results on the CoNLL 2005 data set. The method labels are training/inference. For example, Local/DP means training with the local model, but inference with the dynamic program. Bold font indicates the best system using a single model and a single parse, while the best scores among all systems are underlined. Statistical significance was assessed for F1 and Comp. on the WSJ and Brown test sets with p < 0.01 * and p < 0.05 * * .",
                "num": null,
                "html": null
            },
            "TABREF5": {
                "type_str": "table",
                "content": "<table><tr><td>: Semantic role labeling results on the OntoNotes 5.0 development and test sets from CoNLL 2012. \"Prad-han\" is the Overall results from Table 5 of Pradhan et al. (2013). \"Pradhan (revised)\" are corrected results from per-sonal communication with Pradhan et al. (see footnote 14 for details). Statistical significance was assessed for F1 and Comp. on the test set with p &lt; 0.01</td></tr></table>",
                "text": "",
                "num": null,
                "html": null
            },
            "TABREF7": {
                "type_str": "table",
                "content": "<table/>",
                "text": "Full structure prediction results (joint frame identification and semantic role labeling performance) for FrameNet. All systems use the WSABIE model from Hermann et al.(2014)for the frame identification step. \"Hermann\" is the Wsabie Embedding results fromTable 3ofHermann et al. (2014). Statistical significance was assessed for F1 and Comp. on the test set with p < 0.01 * , p < 0.05 * * and p < 0.075 * * * .",
                "num": null,
                "html": null
            },
            "TABREF8": {
                "type_str": "table",
                "content": "<table/>",
                "text": "Speed and constraint violation results on the OntoNotes 5.0 development set. Exact and approximate methods are shown above and below the line, respectively.",
                "num": null,
                "html": null
            }
        }
    }
}