| { | |
| "paper_id": "R13-1020", | |
| "header": { | |
| "generated_with": "S2ORC 1.0.0", | |
| "date_generated": "2023-01-19T14:55:46.988561Z" | |
| }, | |
| "title": "A Pilot Study on the Semantic Classification of Two German Prepositions: Combining Monolingual and Multilingual Evidence", | |
| "authors": [ | |
| { | |
| "first": "Simon", | |
| "middle": [], | |
| "last": "Clematide", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "", | |
| "institution": "Computational Linguistics University of Zurich", | |
| "location": {} | |
| }, | |
| "email": "[email protected]" | |
| }, | |
| { | |
| "first": "Manfred", | |
| "middle": [], | |
| "last": "Klenner", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "", | |
| "institution": "University of Zurich", | |
| "location": {} | |
| }, | |
| "email": "[email protected]" | |
| } | |
| ], | |
| "year": "", | |
| "venue": null, | |
| "identifiers": {}, | |
| "abstract": "This paper reports on the annotation and maximum-entropy modeling of the semantics of two German prepositions, mit ('with') and auf ('on'). 500 occurrences of each preposition were sampled from a treebank and annotated with syntactosemantic classes by two annotators. The classification is guided by a perspective of information extraction, relies on linguistic tests and aims at the separation of semantically transparent and opaque meanings (that is of collocational constructions). Apart from descriptive statistical material, we present results of experiments using monolingual and multilingual evidence (the latter from informative English and Spanish translations) in order to predict the semantic classes.", | |
| "pdf_parse": { | |
| "paper_id": "R13-1020", | |
| "_pdf_hash": "", | |
| "abstract": [ | |
| { | |
| "text": "This paper reports on the annotation and maximum-entropy modeling of the semantics of two German prepositions, mit ('with') and auf ('on'). 500 occurrences of each preposition were sampled from a treebank and annotated with syntactosemantic classes by two annotators. The classification is guided by a perspective of information extraction, relies on linguistic tests and aims at the separation of semantically transparent and opaque meanings (that is of collocational constructions). Apart from descriptive statistical material, we present results of experiments using monolingual and multilingual evidence (the latter from informative English and Spanish translations) in order to predict the semantic classes.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Abstract", | |
| "sec_num": null | |
| } | |
| ], | |
| "body_text": [ | |
| { | |
| "text": "In linguistics, scientific grammars (Zifonun et al., 1997) as well as grammars for language learners (Helbig and Buscha, 2001 ) follow a long-standing tradition of semantic classification of prepositional phrases. However, it is less well-known which classification schemes can be used for automatic sense disambiguation, supporting for instance applications of information extraction and knowledge discovery.", | |
| "cite_spans": [ | |
| { | |
| "start": 36, | |
| "end": 58, | |
| "text": "(Zifonun et al., 1997)", | |
| "ref_id": "BIBREF25" | |
| }, | |
| { | |
| "start": 101, | |
| "end": 125, | |
| "text": "(Helbig and Buscha, 2001", | |
| "ref_id": null | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Introduction", | |
| "sec_num": "1" | |
| }, | |
| { | |
| "text": "In this pilot study, we want to gain experience of how to classify the semantic contributions of various prepositions from a multilingual perspective. Our main goal is to distinguish between semantically transparent contributions that prepositions can provide in a general or productive manner and the less transparent contributions in collocational constructions. Many prepositions are subcategorized by verbs (or adjectives) and the semantic contribution of a selected preposition is weak or unspecific-a fact that is often revealed by crosslingual comparisons of subcategorization frames. In this study we want to assess the influence of syntactic dependencies and subcategorization on semantic classification. Therefore, we chose to take our material from a syntactically annotated treebank.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Introduction", | |
| "sec_num": "1" | |
| }, | |
| { | |
| "text": "The rest of this paper is organized as follows. Section 2 presents related work and approaches. In Section 3, we describe our syntacto-semantic classification system used in the annotation of prepositions sampled from a German treebank. We also present the types of evidence used in the machine learning experiments for the automatic prediction of the classes. Section 4 contains a systematic evaluation of the performance of the different evidence that we have integrated in our approach.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Introduction", | |
| "sec_num": "1" | |
| }, | |
| { | |
| "text": "As Baldwin et al. (2009, p.134) have put it in their introduction to a special issue on that topic in the Computational Linguistics Journal: \"Information extraction is one application where prepositions are uncontroversially crucial to system accuracy\". The underlying task can be cast as preposition (word) sense disambiguation (WSD). It also has been recognized in the machine translation community that \"prepositions are hard to translate\" (Shilon et al., 2012, p.106) . Although semantic information helps to tackle the translation task, the semantic class of a preposition does not perfectly determine the correct translation. As a consequence, these approaches do not strive to carry out preposition WSD, but to use semantic features in order to more directly map source prepositions to target prepositions (Li et al., 2005) , be it rulebased (Agirre et al., 2009) or with machine learning given aligned bilingual data (Gustavii, 2005) .", | |
| "cite_spans": [ | |
| { | |
| "start": 3, | |
| "end": 31, | |
| "text": "Baldwin et al. (2009, p.134)", | |
| "ref_id": null | |
| }, | |
| { | |
| "start": 443, | |
| "end": 471, | |
| "text": "(Shilon et al., 2012, p.106)", | |
| "ref_id": null | |
| }, | |
| { | |
| "start": 813, | |
| "end": 830, | |
| "text": "(Li et al., 2005)", | |
| "ref_id": "BIBREF14" | |
| }, | |
| { | |
| "start": 849, | |
| "end": 870, | |
| "text": "(Agirre et al., 2009)", | |
| "ref_id": "BIBREF0" | |
| }, | |
| { | |
| "start": 925, | |
| "end": 941, | |
| "text": "(Gustavii, 2005)", | |
| "ref_id": "BIBREF7" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Related Work", | |
| "sec_num": "2" | |
| }, | |
| { | |
| "text": "A great deal of work on preposition classification and WSD has been carried out on the English language. Most prominent the Preposition Project (Litkowski and Hargraves, 2006) that uses a fine-grained classification scheme derived from the Oxford Dictionary (see also the SemEval Task on WSD of prepositions, Litkowski (2007)). Other elaborated classification schemes can be found as part of VerbNet (Kipper et al., 2004) and PrepNet (Saint-Dizier, 2008) .", | |
| "cite_spans": [ | |
| { | |
| "start": 144, | |
| "end": 175, | |
| "text": "(Litkowski and Hargraves, 2006)", | |
| "ref_id": "BIBREF15" | |
| }, | |
| { | |
| "start": 400, | |
| "end": 421, | |
| "text": "(Kipper et al., 2004)", | |
| "ref_id": "BIBREF12" | |
| }, | |
| { | |
| "start": 434, | |
| "end": 454, | |
| "text": "(Saint-Dizier, 2008)", | |
| "ref_id": "BIBREF20" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Related Work", | |
| "sec_num": "2" | |
| }, | |
| { | |
| "text": "Annotated data is available from the Penn Treebank II (Marcus et al., 1994) , where thematic roles occurring with prepositional phrases are marked, and FrameNet (Baker et al., 1998) , which was annotated as part of the Preposition Project. There have been a couple of ACL-SIGSEM workshops on prepositions (the last one in 2007) covering all aspects of preposition processing (not only the semantics).", | |
| "cite_spans": [ | |
| { | |
| "start": 54, | |
| "end": 75, | |
| "text": "(Marcus et al., 1994)", | |
| "ref_id": "BIBREF17" | |
| }, | |
| { | |
| "start": 161, | |
| "end": 181, | |
| "text": "(Baker et al., 1998)", | |
| "ref_id": "BIBREF1" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Related Work", | |
| "sec_num": "2" | |
| }, | |
| { | |
| "text": "On the methodological side, preposition disambiguation sometimes is coupled with semantic role resources, e.g. O'Hara and Wiebe (2009). There, traditional features for WSD (e.g. the preposition, stem of embedded noun, POS and stem of words in a fixed window around the preposition) are augmented with semantic features stemming from knowledge resources such as FrameNet and WordNet (Fellbaum, 1998) . In O'Hara and Wiebe (2009), a new feature, hypernym collocation (the WordNet hypernym of the embedded noun), is used to carry out disambiguation relative to either coarse-grained Penn treebank functional roles or more sophisticated FrameNet roles. Syntactic information, e.g. the syntactic function of the PP, is ignored in their system (in contrast to our approach).", | |
| "cite_spans": [ | |
| { | |
| "start": 382, | |
| "end": 398, | |
| "text": "(Fellbaum, 1998)", | |
| "ref_id": "BIBREF6" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Related Work", | |
| "sec_num": "2" | |
| }, | |
| { | |
| "text": "As can be seen from the discussion above, there is no canonical classification scheme for preposition disambiguation. Furthermore, the semantic class that a preposition can take is language specific. For German, there are but a few approaches (Hartrumpf et al., 2006; M\u00fcller et al., 2011) . M\u00fcller et al. (2011) rely on an annotation scheme derived from various traditional linguistic theories. 22 prepositions are modeled on the basis of 27 top-level senses. A sense hierarchy is defined (especially for temporal and spatial senses) in order to allow for a more flexible and fine-grained classification. Manually specified decision trees are then used to produce the gold standard classifications. This scheme is, for our purposes, far too finegrained and also hard to automatically model by machine learning. However, if their resources sem\\syn opp mod vmod ?mod -p verbal 131 2 3 3 139 nominal 2 120 2 2 3 129 coll 42 3 8 2 55 TEM 6 6 MOD 3 8 4 1 16 LOC 8 10 77 8 5 2 Hartrumpf et al. (2006) is geared towards a semantic formalism called MultiNet (Helbig, 2006) , it fully relies on this proprietary resource.", | |
| "cite_spans": [ | |
| { | |
| "start": 243, | |
| "end": 267, | |
| "text": "(Hartrumpf et al., 2006;", | |
| "ref_id": "BIBREF8" | |
| }, | |
| { | |
| "start": 268, | |
| "end": 288, | |
| "text": "M\u00fcller et al., 2011)", | |
| "ref_id": "BIBREF18" | |
| }, | |
| { | |
| "start": 291, | |
| "end": 311, | |
| "text": "M\u00fcller et al. (2011)", | |
| "ref_id": "BIBREF18" | |
| }, | |
| { | |
| "start": 1006, | |
| "end": 1029, | |
| "text": "Hartrumpf et al. (2006)", | |
| "ref_id": "BIBREF8" | |
| }, | |
| { | |
| "start": 1085, | |
| "end": 1099, | |
| "text": "(Helbig, 2006)", | |
| "ref_id": "BIBREF11" | |
| } | |
| ], | |
| "ref_spans": [ | |
| { | |
| "start": 865, | |
| "end": 1005, | |
| "text": "-p verbal 131 2 3 3 139 nominal 2 120 2 2 3 129 coll 42 3 8 2 55 TEM 6 6 MOD 3 8 4 1 16 LOC 8 10 77 8 5 2", | |
| "ref_id": "TABREF1" | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Related Work", | |
| "sec_num": "2" | |
| }, | |
| { | |
| "text": "As mentioned in Section 2, the Penn Treebank comprises shallow semantic annotations to prepositional phrases (PP). There, a distinction is made between six semantic classes of PPs (and, thus, prepositions): locative, direction, manner, pur-pose, temporal, and extent. Unfortunately, none of the large German treebanks (TIGER (Brants and Hansen, 2002) , T\u00fcba-D/Z (Telljohann et al., 2004) ) provide such a comparable rudimentary scheme that could be a starting point for our pilot study. There is no resource, we could use, although one is currently being developed by another group (M\u00fcller et al., 2011) , but it is not yet released. Since we believe that treebanks could benefit from such an additional annotation layer, we decided to work with a German treebank, the T\u00fcbinger Baumbank T\u00fcba-D/Z 7.0. It comprises about 65,000 annotated sentences, besides phrase structure, also topological fields and grammatical functions are specified. PPs can act as obligatory or optional (opp) complements of verbs, or as adjuncts (vmod). In the current study, we mainly focus on PPs acting as verb complements (opp) or adjuncts (vmod) .", | |
| "cite_spans": [ | |
| { | |
| "start": 325, | |
| "end": 350, | |
| "text": "(Brants and Hansen, 2002)", | |
| "ref_id": "BIBREF3" | |
| }, | |
| { | |
| "start": 362, | |
| "end": 387, | |
| "text": "(Telljohann et al., 2004)", | |
| "ref_id": "BIBREF23" | |
| }, | |
| { | |
| "start": 582, | |
| "end": 603, | |
| "text": "(M\u00fcller et al., 2011)", | |
| "ref_id": "BIBREF18" | |
| }, | |
| { | |
| "start": 1118, | |
| "end": 1124, | |
| "text": "(vmod)", | |
| "ref_id": null | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Resources", | |
| "sec_num": "3.1" | |
| }, | |
| { | |
| "text": "From the ten most frequent prepositions in the T\u00fcba-D/Z we have chosen one with a predominant local and temporal meaning (auf 'on') and one with more broader meaning spectrum (mit 'with'). We randomly sampled 500 occurrences of each preposition from the T\u00fcba-D/Z and annotated each preposition according to our classification scheme described below.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Resources", | |
| "sec_num": "3.1" | |
| }, | |
| { | |
| "text": "The intended application is information extraction and question answering. Accordingly, our semantic classes had to be tightly coupled with question words. That is, the way users may ask, determines the granularity of the classification scheme. Typical interrogative words and phrases are how (modal), how long (temporal, duration), when (temporal, time point), where (locative).", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Semantics of auf and mit", | |
| "sec_num": "3.1.1" | |
| }, | |
| { | |
| "text": "In the case of auf (cf. Table 1 ), we distinguish between locative (LOC where), directional (DIR where to), temporal (TEM when, how long), modal (MOD how), and causal (CAUS why) PPs. If in a temporal PP the noun is an event (e.g. party), then often a locative or a temporal reading is possible (e.g. when or where did he laugh? -at his party). We use TLOC to refer to this usage. If the PP acts as a modifier of an adjective or noun, it is annotated with \"nominal\" (e.g. 'the cup on the table'). For the preposition auf, we have annotated currently only adjuncts and verb complements with their semantic classes. In case that the verb governs an otherwise semantically vacuous preposition (warten auf 'to wait for'), the preposition is marked with \"verbal\". Finally, any idiomatic expression comprising a PP having a non-compositional meaning like auf den Putz hauen, 'to kick up one's heels' is annotated as collocational (\"coll\"). The preposition does not contribute any semantics in these cases. Sometimes no decision was possible (e.g. given sentence fragments, missing global context, unclear semantics), we used \"?\" to annotate these cases. Table 1 shows the distribution of these classes and their syntactic realization. Verb/preposition collocations form the largest class (139), followed by nominal modification (129) and locatives (110). Syntactically, there are three groups to be distinguished: PP complements (opp, 202), NP and PP modification (mod, 136) and adjuncts (vmod, 125) . The table reveals a moderate number of interpretation divergences between the T\u00fcba-D/Z annotators and us. Some stem from structural ambiguity (e.g. \"?mod\" denotes PP attachment ambiguities), and are to be expected. Ideally, however, if a PP bears the functional label \"mod\", it should be classified as \"nominal\" in our scheme. Also, a \"vmod\" should not be annotated as \"verbal\", since \"verbal\" means that the preposition is vacuous, while \"vmod\" means that it acts as an adjunct. For instance, we disagreed with 3 \"vmod\" (adjuncts) and interpreted them as verbpreposition collocations, also 2 \"vmod\" are better classified as \"nominal\" in our view. However, the majority of decisions does not contradict or even is in line with the functional assignments of the T\u00fcba-D/Z. For example, of the 136 \"mod\" (NP or PP modifications), we placed 120 in our corresponding class \"nominal\".", | |
| "cite_spans": [ | |
| { | |
| "start": 1481, | |
| "end": 1492, | |
| "text": "(vmod, 125)", | |
| "ref_id": null | |
| } | |
| ], | |
| "ref_spans": [ | |
| { | |
| "start": 24, | |
| "end": 31, | |
| "text": "Table 1", | |
| "ref_id": "TABREF1" | |
| }, | |
| { | |
| "start": 1147, | |
| "end": 1154, | |
| "text": "Table 1", | |
| "ref_id": "TABREF1" | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Semantics of auf and mit", | |
| "sec_num": "3.1.1" | |
| }, | |
| { | |
| "text": "In the case of mit (cf. Table 2), the syntactic classification labels \"verb\", \"nominal\" and \"coll\" are used as introduced above for auf. The prepositions auf and mit also share two core semantic classes, namely TEM (temporal) and MOD (modal). The other semantic classes of mit are: COM for comitative use (to watch a movie with a friend), ORN for ornative use (to tell with humor), SIZ indicating size or extent (to demonstrate with 100 people against), INS for the instrument reading which is a subclass of MOD (modal) (to break with a hammer), and IDE for identity (with him, hope enters the room meaning: he represents/is identical with hope).", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Semantics of auf and mit", | |
| "sec_num": "3.1.1" | |
| }, | |
| { | |
| "text": "As with auf, there are some divergences between the functional annotations of the T\u00fcba-D/Z and our annotation decisions, especially concerning \"vmod\" and \"mod\". We have not fully traced these divergences back to their origins, but see the previous discussion in the context of auf.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Semantics of auf and mit", | |
| "sec_num": "3.1.1" | |
| }, | |
| { | |
| "text": "We have measured inter-annotator agreement in two stages: after our initial annotation round, and after some discussion and refinements of our annotation scheme in a second step on a harmonized version of the data. One reason for disagreement concerning mit was the annotation with \"ornative\": a rather sophisticated annotation scheme would allow the use of ORN even in cases where it is modal, but also implicitely qualifies the subject of the sentence (he says it with a gentle voice). In these cases, however, it is more natural to ask how (has he said it), so we disallowed ORN in such examples.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Inter-Annotator Agreement", | |
| "sec_num": "3.1.2" | |
| }, | |
| { | |
| "text": "We report the annotator agreement as percentage of agreeing pairs and as Cohen's \u03ba. The initial inter-annotator agreement for mit was 85% (\u03ba = .82), while after harmonization it was 91.8% (\u03ba = .90) and 92% (\u03ba = .90) between the harmonized version and the separately created initial annotations of the two annotators, respectively. With auf the agreement was lower, namely initially 74% (\u03ba = .67). After harmonization is was 84.8% (\u03ba = .81) and 86.2% (\u03ba = .83), respectively. The main source of confusion here was the annotation scheme of PPs in the context of local verbs (LOC and DIR). The question was whether to treat these roles as adjuncts or as verb complements. Also the decision when to treat a verb-preposition combination as a collocation or not, was not sufficiently well described and operationalized in the guidelines.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Inter-Annotator Agreement", | |
| "sec_num": "3.1.2" | |
| }, | |
| { | |
| "text": "The rationale behind our two-stage procedure was to first independently create annotation strategies based on existing classes from the literature and to later refine them to valid annotation guidelines based on the evidence found in the data.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Inter-Annotator Agreement", | |
| "sec_num": "3.1.2" | |
| }, | |
| { | |
| "text": "As already mentioned, prepositional semantics is language-specific: The semantic classes a preposition might express do vary between languages, the semantic contributions given by a preposition in one language are often realized by different prepositions in different languages. Moreover, the semantic functions a preposition (e.g. mit) and its direct translation ('with') can bear, might differ.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Multilingual Evidence", | |
| "sec_num": "3.1.3" | |
| }, | |
| { | |
| "text": "The identity reading of German mit is not possible for English 'with'.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Multilingual Evidence", | |
| "sec_num": "3.1.3" | |
| }, | |
| { | |
| "text": "The question is, whether a multilingual perspective (for instance in the form of Statistical Machine Translation (SMT)) helps determining the semantic class of a given preposition in the source language. Tables 3 and 4 give a detailed overview of how the prepositions mit and auf are translated into English and Spanish by Google Translate. 1 For instance, mit is translated into English as with, of, to, by, in, or not at all (\"0\"). Of course, there are predominant translations, for instance mit was translated 372 times by with and con. There is also a tendency to choose equivalent prepositions across languages, e.g. a and to (Table 4 : 71 cases), but quite often different prepositions are selected. Since we use imperfect translations from SMT we cannot be sure whether the aforementioned differences stem from mistranslations or whether they reveal a true difference. In order to clarify this question one could exploit parallel treebanks. However, currently available resources covering German such as SMULTRON (Volk et al., 2010) still have a limited size (approx. 2500 sentences).", | |
| "cite_spans": [ | |
| { | |
| "start": 1021, | |
| "end": 1040, | |
| "text": "(Volk et al., 2010)", | |
| "ref_id": "BIBREF24" | |
| } | |
| ], | |
| "ref_spans": [ | |
| { | |
| "start": 204, | |
| "end": 218, | |
| "text": "Tables 3 and 4", | |
| "ref_id": "TABREF4" | |
| }, | |
| { | |
| "start": 631, | |
| "end": 640, | |
| "text": "(Table 4", | |
| "ref_id": "TABREF5" | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Multilingual Evidence", | |
| "sec_num": "3.1.3" | |
| }, | |
| { | |
| "text": "The question is whether inter-language divergence of preposition usage helps to determine the semantic class of a preposition in the source language. Or more technically, whether there is a correlation between semantic classes of the source language preposition and a translation made by SMT. Even if such a correlation turns out not to be a strong one, it might nevertheless help as a feature in a machine learning model.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Multilingual Evidence", | |
| "sec_num": "3.1.3" | |
| }, | |
| { | |
| "text": "For illustration purposes, we give two examples of semantic annotations of PPs and the mapping of the German prepositions therein to English prepositions via automatic translation with Google translate.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Annotation and Translation: Examples", | |
| "sec_num": "3.1.4" | |
| }, | |
| { | |
| "text": "In the first case, auf does not carry any semantics, it is part of the verb (warten auf ). Accordingly, it is annotated as \"verbal\". In English, the corresponding verb construction is 'to listen to', es\\en with 0 by to of in on about as from con 372 7 1 6 1 1 388 0 3 48 2 1 1 1 1 57 de 10 5 1 6 1 23 a 7 5 1 1 14 por 1 1 6 8 en 1 2 2 The second examples illustrates that the same semantic class, LOC (local), might be realized by two different prepositions in German and in English. The preposition auf in German can be used to indicate the 'place of living' of a person, if it is a small island (like Sardinia). This is not possible in English. The sentence pairs are: Selbst wenn sie in entlegenen St\u00e4dtchen auf Sardinien leben and 'Even if they live in remote town in Sardinia'.", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 242, | |
| "end": 372, | |
| "text": "con 372 7 1 6 1 1 388 0 3 48 2 1 1 1 1 57 de 10 5 1 6 1 23 a 7 5 1 1 14 por 1 1 6 8 en 1 2 2", | |
| "ref_id": "TABREF1" | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Annotation and Translation: Examples", | |
| "sec_num": "3.1.4" | |
| }, | |
| { | |
| "text": "Note that in these examples auf was not mapped to its direct translation which is 'on'.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Annotation and Translation: Examples", | |
| "sec_num": "3.1.4" | |
| }, | |
| { | |
| "text": "In order to measure the difficulty of an automatic classification of the syntacto-semantic classes expressed by auf and mit we conducted several experiments with the Maximum-Entropy Modeling tool MegaM (Daum\u00e9 III, 2008 ). 2 For this pilot study, we focused on simple features gained from the syntactical configuration (perfect data from the T\u00fcba-D/Z), textual data from the context, and multilingual evidence from Spanish and English translations (imperfect Google translations).", | |
| "cite_spans": [ | |
| { | |
| "start": 202, | |
| "end": 218, | |
| "text": "(Daum\u00e9 III, 2008", | |
| "ref_id": "BIBREF4" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Supervised Machine Learning Approach", | |
| "sec_num": "3.2" | |
| }, | |
| { | |
| "text": "In Section 4 we present and analyze the contribution of the following feature sets:", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Supervised Machine Learning Approach", | |
| "sec_num": "3.2" | |
| }, | |
| { | |
| "text": "head Word, part of speech (POS), and lemma of the head word (typically a noun) of the dependent phrase of the preposition, for instance, the head of mit Sorgfalt is Sorgfalt 'care'. In case of coordinated PPs and multi-word heads, the first token was selected. syntax The syntactic function of the PP taken from the T\u00fcbaD/Z. neighbor Word, POS, and lemma of the preceding and following token. context Word, POS, and lemma in a window of 5 preceding and following tokens (taken as a bag of words, lemmas and POS). en English translation of the preposition produced by the Google translation of the German sentence. es Spanish Google translation of the preposition.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Supervised Machine Learning Approach", | |
| "sec_num": "3.2" | |
| }, | |
| { | |
| "text": "The evaluations described below assess the performance improvement for the multi-class predictions of our annotated prepositions (500 occurrences each) by using different sets of features as evidence. We evaluate against a baseline system which basically predicts the majority class given the lack of any additional evidence. All results are reported as mean accuracy computed by crossvalidation. No stratification of class labels has been applied to the folds of the cross-validation. Accuracy is the proportion of true classifications delivered by the system.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Results and Discussion", | |
| "sec_num": "4" | |
| }, | |
| { | |
| "text": "We performed a 10-fold cross-validation evaluation for the scenario of predicting the full set of all syntactic and semantic classes (cf. Table1 and 2).", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Syntacto-Semantic Classification", | |
| "sec_num": "4.1" | |
| }, | |
| { | |
| "text": "The evaluation results of auf are shown in Table 5. The best system uses the feature sets \"head\", \"neighbor\" and \"syntax\", however, \"syntax\" is by far the strongest feature. If perfect syntactic analyses are not available, \"head\" and \"neighbor\" information can compensate for more than 2/3 of the performance gain. The effect of \"syntax\" is especially strong for auf because the nominal modifiers are classified according to syntactic criteria only. A future, more semantically oriented classification of noun modifiers will probably weaken this effect. Multilingual evidence from informative Google translations improves considerably over the (weak) baseline. Combining the evidence from Spanish and English performs slightly better than each language separately does. Therefore, translations into multiple languages are useful for the case of auf. However, the best systems are those without any translation evidence from Spanish or English. Table 6 shows the corresponding results for mit. The overall performance is lower but the feature sets have a very similar ranking of predictive power. The lower performance stems from the fact that mit has 11 syntacto-semantic classes with a more uniform distribution than auf (10 classes). The best system without the feature \"syntax\" involves 3 different feature sets, \"context\", \"neighbor\" and \"en\". However, these feature sets can only compensate for less than half of the performance gain of the feature set \"syntax\" derived from the treebank syntax structure. The best system performance is reached if either English or Table 5 : Performance of feature sets for syntactosemantic classification accuracy: auf (N = 500).", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 944, | |
| "end": 951, | |
| "text": "Table 6", | |
| "ref_id": null | |
| }, | |
| { | |
| "start": 1571, | |
| "end": 1578, | |
| "text": "Table 5", | |
| "ref_id": null | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Syntacto-Semantic Classification", | |
| "sec_num": "4.1" | |
| }, | |
| { | |
| "text": "The column \"Mean\" contains the average accuracy computed from the cross-validation sets. The column \u2206 abs bs contains the absolute performance gain with respect to the baseline. \u2206 rel bs expresses the relative performance gain. The last row contains the feature set with the best performance. Table 6 : Performance of feature sets for syntactosemantic classification accuracy: mit (N = 500).", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 293, | |
| "end": 300, | |
| "text": "Table 6", | |
| "ref_id": null | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Syntacto-Semantic Classification", | |
| "sec_num": "4.1" | |
| }, | |
| { | |
| "text": "Spanish evidence is added. However, the improvement given by multilingual evidence is rather small.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Syntacto-Semantic Classification", | |
| "sec_num": "4.1" | |
| }, | |
| { | |
| "text": "In a further evaluation, we measured how well the purely semantic classes (i.e. those without \"nominal\", \"verb\" and \"coll\") can be predicted. For auf we have 171 cases with a defined semantic classification, for mit 290. Due to the smaller training sizes we performed 5-fold cross-validation. Table 7 illustrates the problems from the skewed distribution of semantic classes in the case of auf : Just guessing the largest class LOC represents a baseline decision which is hard to beat. Only the feature set \"head\" can improve over this baseline, all other features either deteriorate the system performance or do not improve it. Interestingly, the best system combines the translation evidence from Spanish with the feature set \"head\". Adding more feature sets does not improve the results (see Table 7 second last row). The less skewed distribution of semantic classes in the case of mit allows for a significant improvement over the baseline system. Table 8 shows that most feature sets have a beneficial effect, and therefore, classification performance is almost doubled by the best system. In contrast to the syntacto-semantic classification, multilingual evidence does not contribute to the best system. The only configuration where multilingual evidence improves performance appears if the syntactic dependency information from the treebank is dropped. The best system without the feature set \"syntax\" relies on English and Spanish evidence.", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 293, | |
| "end": 300, | |
| "text": "Table 7", | |
| "ref_id": "TABREF9" | |
| }, | |
| { | |
| "start": 795, | |
| "end": 802, | |
| "text": "Table 7", | |
| "ref_id": "TABREF9" | |
| }, | |
| { | |
| "start": 952, | |
| "end": 959, | |
| "text": "Table 8", | |
| "ref_id": "TABREF10" | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Semantic Classification", | |
| "sec_num": "4.2" | |
| }, | |
| { | |
| "text": "The results of our experiments in using multilingual evidence for the syntacto-semantic and semantic classification of prepositions are mixed. The syntactico-semantic classification of auf works best without multilingual evidence although there is a weak correlation between the feature sets \"en\" and \"es\" and the syntacto-semantic classes. However, the best system of the syntacticosemantic classification of mit profits from added multilingual evidence although this evidence taken as a single feature set cannot beat the baseline.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Semantic Classification", | |
| "sec_num": "4.2" | |
| }, | |
| { | |
| "text": "For the purely semantic classification, no improvement over the baseline can be found by the multilingual evidence for both prepositions. Still, multilingual evidence helps in these cases where syntactic information is not valuable (in the case of auf ), or if we mute this feature (in the case of mit).", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Semantic Classification", | |
| "sec_num": "4.2" | |
| }, | |
| { | |
| "text": "Our annotation and modeling experiments illustrate the different semantic and distributional characteristics of the considered German prepositions auf and mit. The skewed distribution of the semantic classes of auf represent a challenge for any classifier. If small semantic classes should be detected, more training material is needed for these cases. The application of Active Learning techniques (Settles, 2012) might help to efficiently collect such data.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Conclusion", | |
| "sec_num": "5" | |
| }, | |
| { | |
| "text": "Our experiments with maximum entropy modeling indicate that informative Google translations of prepositions do not lead to a significant performance improvement in semantic classification. Simple monolingual contextual features generally perform better. The inclusion of perfect (i.e. treebank-derived) syntactic dependency information generally performs best. However, for practical systems only imperfect syntax analyses from error-producing parsers are available. Future research is needed to assess the performance decrease if parser output is provided instead of handcrafted manual annotation.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Conclusion", | |
| "sec_num": "5" | |
| }, | |
| { | |
| "text": "Another topic for future work is the integration of further language resources. Bilingual lexicons such as dict.cc 3 contain information about semantically void subcategorized prepositions, for instance auf jdn warten is linked to to wait for sb. Statistical collocation analyses derived from large German corpora are provided by services such as \"Wortschatz Leipzig\" 4 or \"Digitales W\u00f6rterbuch der Deutschen Sprache\" 5 .", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Conclusion", | |
| "sec_num": "5" | |
| }, | |
| { | |
| "text": "Given the available amount of electronic texts, the application of distributional semantics for preposition disambiguation and for modeling of the semantic fingerprint of prepositions also seems promising (cf. (de Cruys and Apidianaki, 2011)).", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Conclusion", | |
| "sec_num": "5" | |
| }, | |
| { | |
| "text": "Finally, contextual features might profit from synonym expansion or synonym set classification, a technique also used by Kiss et al. (2010) .", | |
| "cite_spans": [ | |
| { | |
| "start": 121, | |
| "end": 139, | |
| "text": "Kiss et al. (2010)", | |
| "ref_id": "BIBREF13" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Conclusion", | |
| "sec_num": "5" | |
| }, | |
| { | |
| "text": "For this experiment, we manually mapped the prepositions from the translated sentences using the phrase alignment visualization of http://translate.google. com. English and Spanish was chosen since according to http://matrix.statmt.org/matrix the translation quality of German to English and Spanish is best and at the same time both target languages belong to different language families.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "Maximum-Entropy modeling is also known as logistic regression. In our experiments, we used the default regularization parameter \u03bb = 1 of MegaM.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "See http://www.dict.cc 4 See http://wortschatz.uni-leipzig.de 5 See http://dwds.de", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "", | |
| "sec_num": null | |
| } | |
| ], | |
| "back_matter": [], | |
| "bib_entries": { | |
| "BIBREF0": { | |
| "ref_id": "b0", | |
| "title": "Use of rich linguistic information to translate prepositions and grammar cases to Basque", | |
| "authors": [ | |
| { | |
| "first": "Eneko", | |
| "middle": [], | |
| "last": "Agirre", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Aitziber", | |
| "middle": [], | |
| "last": "Atutxa", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Gorka", | |
| "middle": [], | |
| "last": "Labak", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Mikel", | |
| "middle": [], | |
| "last": "Lersundi", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2009, | |
| "venue": "Proceedings of the XIII Conference of the European Association for Machine Translation (EAMT)", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "58--65", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Eneko Agirre, Aitziber Atutxa, Gorka Labak, Mikel Lersundi, Aingeru Mayor, and Kepa Sarasola. 2009. Use of rich linguistic information to translate prepo- sitions and grammar cases to Basque. In Proceed- ings of the XIII Conference of the European Associa- tion for Machine Translation (EAMT), pages 58-65.", | |
| "links": null | |
| }, | |
| "BIBREF1": { | |
| "ref_id": "b1", | |
| "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": "COLING-ACL", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "86--90", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Collin F. Baker, Charles J. Fillmore, and John B. Lowe. 1998. The Berkeley FrameNet project. In Christian Boitet and Pete Whitelock, editors, COLING-ACL, pages 86-90. Morgan Kaufmann Publishers / ACL.", | |
| "links": null | |
| }, | |
| "BIBREF2": { | |
| "ref_id": "b2", | |
| "title": "Prepositions in applications: A survey and introduction to the special issue", | |
| "authors": [ | |
| { | |
| "first": "Timothy", | |
| "middle": [], | |
| "last": "Baldwin", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Valia", | |
| "middle": [], | |
| "last": "Kordoni", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Aline", | |
| "middle": [], | |
| "last": "Villavicencio", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2009, | |
| "venue": "Computational Linguistics", | |
| "volume": "35", | |
| "issue": "2", | |
| "pages": "119--149", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Timothy Baldwin, Valia Kordoni, and Aline Villavicen- cio. 2009. Prepositions in applications: A survey and introduction to the special issue. Computational Linguistics, 35(2):119-149.", | |
| "links": null | |
| }, | |
| "BIBREF3": { | |
| "ref_id": "b3", | |
| "title": "Developments in the TIGER annotation scheme and their realization in the corpus", | |
| "authors": [ | |
| { | |
| "first": "Sabine", | |
| "middle": [], | |
| "last": "Brants", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Silvia", | |
| "middle": [], | |
| "last": "Hansen", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2002, | |
| "venue": "Proceedings of the Third Conference on Language Resources and Evaluation (LREC 2002)", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "1643--1649", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Sabine Brants and Silvia Hansen. 2002. Develop- ments in the TIGER annotation scheme and their re- alization in the corpus. In Proceedings of the Third Conference on Language Resources and Evaluation (LREC 2002), pages 1643-1649, Las Palmas.", | |
| "links": null | |
| }, | |
| "BIBREF4": { | |
| "ref_id": "b4", | |
| "title": "MegaM: Maximum entropy model optimization package. ACL Data and Code Repository, ADCR2008C003", | |
| "authors": [ | |
| { | |
| "first": "Hal", | |
| "middle": [], | |
| "last": "Daum\u00e9", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Iii", | |
| "middle": [], | |
| "last": "", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2008, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Hal Daum\u00e9 III. 2008. MegaM: Maximum entropy model optimization package. ACL Data and Code Repository, ADCR2008C003.", | |
| "links": null | |
| }, | |
| "BIBREF5": { | |
| "ref_id": "b5", | |
| "title": "Latent semantic word sense induction and disambiguation", | |
| "authors": [ | |
| { | |
| "first": "Tim", | |
| "middle": [], | |
| "last": "Van De Cruys", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Marianna", | |
| "middle": [], | |
| "last": "Apidianaki", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2011, | |
| "venue": "ACL", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "1476--1485", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Tim Van de Cruys and Marianna Apidianaki. 2011. Latent semantic word sense induction and disam- biguation. In Dekang Lin, Yuji Matsumoto, and Rada Mihalcea, editors, ACL, pages 1476-1485. The Association for Computer Linguistics.", | |
| "links": null | |
| }, | |
| "BIBREF6": { | |
| "ref_id": "b6", | |
| "title": "WordNet: An Electronic Lexical Database", | |
| "authors": [ | |
| { | |
| "first": "Christiane", | |
| "middle": [], | |
| "last": "Fellbaum", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 1998, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Christiane Fellbaum. 1998. WordNet: An Electronic Lexical Database. Bradford Books.", | |
| "links": null | |
| }, | |
| "BIBREF7": { | |
| "ref_id": "b7", | |
| "title": "Target language preposition selection -an experiment with transformation-based learning and aligned bilingual data", | |
| "authors": [ | |
| { | |
| "first": "Ebba", | |
| "middle": [], | |
| "last": "Gustavii", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2005, | |
| "venue": "Proceedings of the XIII Conference of the European Association for Machine Translation (EAMT)", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "112--118", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Ebba Gustavii. 2005. Target language preposition se- lection -an experiment with transformation-based learning and aligned bilingual data. In Proceedings of the XIII Conference of the European Association for Machine Translation (EAMT), pages 112-118.", | |
| "links": null | |
| }, | |
| "BIBREF8": { | |
| "ref_id": "b8", | |
| "title": "Semantic interpretation of prepositions for NLP applications", | |
| "authors": [ | |
| { | |
| "first": "Sven", | |
| "middle": [], | |
| "last": "Hartrumpf", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Hermann", | |
| "middle": [], | |
| "last": "Helbig", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Rainer", | |
| "middle": [], | |
| "last": "Osswald", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2006, | |
| "venue": "Proceedings of the Third ACL-SIGSEM Workshop on Prepositions", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "29--36", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Sven Hartrumpf, Hermann Helbig, and Rainer Oss- wald. 2006. Semantic interpretation of prepositions for NLP applications. In Proceedings of the Third ACL-SIGSEM Workshop on Prepositions, pages 29- 36.", | |
| "links": null | |
| }, | |
| "BIBREF10": { | |
| "ref_id": "b10", | |
| "title": "Ein Handbuch f\u00fcr den Ausl\u00e4nderunterricht. Langenscheidt", | |
| "authors": [ | |
| { | |
| "first": "Deutsche", | |
| "middle": [], | |
| "last": "Grammatik", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": null, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Deutsche Grammatik: Ein Handbuch f\u00fcr den Ausl\u00e4nderunterricht. Langenscheidt.", | |
| "links": null | |
| }, | |
| "BIBREF11": { | |
| "ref_id": "b11", | |
| "title": "Knowledge Representation and the Semantics of Natural Language. Cognitive Technologies", | |
| "authors": [ | |
| { | |
| "first": "Hermann", | |
| "middle": [], | |
| "last": "Helbig", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2006, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Hermann Helbig. 2006. Knowledge Representation and the Semantics of Natural Language. Cognitive Technologies. Springer.", | |
| "links": null | |
| }, | |
| "BIBREF12": { | |
| "ref_id": "b12", | |
| "title": "Using prepositions to extend a verb lexicon", | |
| "authors": [ | |
| { | |
| "first": "Karin", | |
| "middle": [], | |
| "last": "Kipper", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Benjamin", | |
| "middle": [], | |
| "last": "Snyder", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Martha", | |
| "middle": [], | |
| "last": "Palmer", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2004, | |
| "venue": "Proceedings of the HLT/NAACL Workshop on Computational Lexical Semantics", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "23--29", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Karin Kipper, Benjamin Snyder, and Martha Palmer. 2004. Using prepositions to extend a verb lexi- con. In Proceedings of the HLT/NAACL Workshop on Computational Lexical Semantics, pages 23-29.", | |
| "links": null | |
| }, | |
| "BIBREF13": { | |
| "ref_id": "b13", | |
| "title": "A logistic regression model of determiner omission in PPs", | |
| "authors": [ | |
| { | |
| "first": "Tibor", | |
| "middle": [], | |
| "last": "Kiss", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Katja", | |
| "middle": [], | |
| "last": "Ke\u00dfelmeier", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Antje", | |
| "middle": [], | |
| "last": "M\u00fcller", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Claudia", | |
| "middle": [], | |
| "last": "Roch", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Tobias", | |
| "middle": [], | |
| "last": "Stadtfeld", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2010, | |
| "venue": "COLING (Posters)", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "561--569", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Tibor Kiss, Katja Ke\u00dfelmeier, Antje M\u00fcller, Claudia Roch, Tobias Stadtfeld, and Jan Strunk. 2010. A logistic regression model of determiner omission in PPs. In Chu-Ren Huang and Dan Jurafsky, editors, COLING (Posters), pages 561-569. Chinese Infor- mation Processing Society of China.", | |
| "links": null | |
| }, | |
| "BIBREF14": { | |
| "ref_id": "b14", | |
| "title": "English to Chinese translation of prepositions", | |
| "authors": [ | |
| { | |
| "first": "Hui", | |
| "middle": [], | |
| "last": "Li", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Nathalie", | |
| "middle": [], | |
| "last": "Japkowicz", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Caroline", | |
| "middle": [], | |
| "last": "Barri\u00e8re", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2005, | |
| "venue": "Canadian Conference on AI", | |
| "volume": "3501", | |
| "issue": "", | |
| "pages": "412--416", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Hui Li, Nathalie Japkowicz, and Caroline Barri\u00e8re. 2005. English to Chinese translation of preposi- tions. In Canadian Conference on AI, volume 3501 of Lecture Notes in Computer Science, pages 412- 416. Springer.", | |
| "links": null | |
| }, | |
| "BIBREF15": { | |
| "ref_id": "b15", | |
| "title": "Coverage and inheritance in The Preposition Project", | |
| "authors": [ | |
| { | |
| "first": "C", | |
| "middle": [], | |
| "last": "Kenneth", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Orin", | |
| "middle": [], | |
| "last": "Litkowski", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "Hargraves", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2006, | |
| "venue": "Third ACL-SIGSEM Workshop on Prepositions", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "37--44", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Kenneth C. Litkowski and Orin Hargraves. 2006. Cov- erage and inheritance in The Preposition Project. In Third ACL-SIGSEM Workshop on Prepositions, pages 37-44.", | |
| "links": null | |
| }, | |
| "BIBREF16": { | |
| "ref_id": "b16", | |
| "title": "CLR: Integration of FrameNet in a text representation system", | |
| "authors": [ | |
| { | |
| "first": "C", | |
| "middle": [], | |
| "last": "Kenneth", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "Litkowski", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2007, | |
| "venue": "Proceedings of the Fourth International Workshop on Semantic Evaluations (SemEval-2007)", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "113--116", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Kenneth C. Litkowski. 2007. CLR: Integration of FrameNet in a text representation system. In Pro- ceedings of the Fourth International Workshop on Semantic Evaluations (SemEval-2007), pages 113- 116, Prague, Czech Republic, June. ACL.", | |
| "links": null | |
| }, | |
| "BIBREF17": { | |
| "ref_id": "b17", | |
| "title": "The Penn treebank: Annotating predicate argument structure", | |
| "authors": [ | |
| { | |
| "first": "Mitchell", | |
| "middle": [ | |
| "P" | |
| ], | |
| "last": "Marcus", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Grace", | |
| "middle": [], | |
| "last": "Kim", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Mary", | |
| "middle": [ | |
| "Ann" | |
| ], | |
| "last": "Marcinkiewicz", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Robert", | |
| "middle": [], | |
| "last": "Macintyre", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Ann", | |
| "middle": [], | |
| "last": "Bies", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Mark", | |
| "middle": [], | |
| "last": "Ferguson", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Karen", | |
| "middle": [], | |
| "last": "Katz", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Britta", | |
| "middle": [], | |
| "last": "Schasberger", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 1994, | |
| "venue": "HLT", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "114--119", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Mitchell P. Marcus, Grace Kim, Mary Ann Marcinkiewicz, Robert MacIntyre, Ann Bies, Mark Ferguson, Karen Katz, and Britta Schas- berger. 1994. The Penn treebank: Annotating predicate argument structure. In HLT, pages 114-119. Morgan Kaufmann.", | |
| "links": null | |
| }, | |
| "BIBREF18": { | |
| "ref_id": "b18", | |
| "title": "Annotating spatial interpretations of German prepositions", | |
| "authors": [ | |
| { | |
| "first": "Antje", | |
| "middle": [], | |
| "last": "M\u00fcller", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Claudia", | |
| "middle": [], | |
| "last": "Roch", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Tobias", | |
| "middle": [], | |
| "last": "Stadtfeld", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Tibor", | |
| "middle": [], | |
| "last": "Kiss", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2011, | |
| "venue": "ICSC", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "459--466", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Antje M\u00fcller, Claudia Roch, Tobias Stadtfeld, and Ti- bor Kiss. 2011. Annotating spatial interpretations of German prepositions. In ICSC, pages 459-466. IEEE.", | |
| "links": null | |
| }, | |
| "BIBREF19": { | |
| "ref_id": "b19", | |
| "title": "Exploiting semantic role resources for preposition disambiguation", | |
| "authors": [ | |
| { | |
| "first": "Janyce", | |
| "middle": [], | |
| "last": "Tom O'hara", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "Wiebe", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2009, | |
| "venue": "Computational Linguistics", | |
| "volume": "35", | |
| "issue": "2", | |
| "pages": "151--184", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Tom O'Hara and Janyce Wiebe. 2009. Exploiting semantic role resources for preposition disambigua- tion. Computational Linguistics, 35(2):151-184.", | |
| "links": null | |
| }, | |
| "BIBREF20": { | |
| "ref_id": "b20", | |
| "title": "Syntactic and semantic frames in PrepNet", | |
| "authors": [ | |
| { | |
| "first": "Patrick", | |
| "middle": [], | |
| "last": "Saint-Dizier", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2008, | |
| "venue": "IJCNLP", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "763--768", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Patrick Saint-Dizier. 2008. Syntactic and semantic frames in PrepNet. In IJCNLP, pages 763-768. ACL.", | |
| "links": null | |
| }, | |
| "BIBREF21": { | |
| "ref_id": "b21", | |
| "title": "Active Learning", | |
| "authors": [], | |
| "year": null, | |
| "venue": "Synthesis Lectures on Artificial Intelligence and Machine Learning. Morgan & Claypool", | |
| "volume": "6", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Burr Settles. 2012. Active Learning, volume 6 of Synthesis Lectures on Artificial Intelligence and Ma- chine Learning. Morgan & Claypool.", | |
| "links": null | |
| }, | |
| "BIBREF22": { | |
| "ref_id": "b22", | |
| "title": "Incorporating linguistic knowledge in statistical machine translation: Translating prepositions", | |
| "authors": [ | |
| { | |
| "first": "Reshef", | |
| "middle": [], | |
| "last": "Shilon", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Hanna", | |
| "middle": [], | |
| "last": "Fadida", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Shuly", | |
| "middle": [], | |
| "last": "Wintner", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2012, | |
| "venue": "Proceedings of the EACL-2012 Workshop on Innovative Hybrid Approaches to the Processing of Textual Data", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "106--114", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Reshef Shilon, Hanna Fadida, and Shuly Wintner. 2012. Incorporating linguistic knowledge in statisti- cal machine translation: Translating prepositions. In Proceedings of the EACL-2012 Workshop on Inno- vative Hybrid Approaches to the Processing of Tex- tual Data, pages 106-114.", | |
| "links": null | |
| }, | |
| "BIBREF23": { | |
| "ref_id": "b23", | |
| "title": "The T\u00fcba-D/Z Treebank: Annotating German with a Context-Free Backbone", | |
| "authors": [ | |
| { | |
| "first": "Heike", | |
| "middle": [], | |
| "last": "Telljohann", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Erhard", | |
| "middle": [ | |
| "W" | |
| ], | |
| "last": "Hinrichs", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Sandra", | |
| "middle": [], | |
| "last": "K\u00fcbler", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2004, | |
| "venue": "LREC", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "2229--2232", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Heike Telljohann, Erhard W. Hinrichs, and Sandra K\u00fcbler. 2004. The T\u00fcba-D/Z Treebank: Anno- tating German with a Context-Free Backbone. In LREC, pages 2229-2232. European Language Re- sources Association.", | |
| "links": null | |
| }, | |
| "BIBREF24": { | |
| "ref_id": "b24", | |
| "title": "SMULTRON (version 3.0) -The Stockholm MULtilingual parallel TReebank", | |
| "authors": [ | |
| { | |
| "first": "Martin", | |
| "middle": [], | |
| "last": "Volk", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Anne", | |
| "middle": [], | |
| "last": "G\u00f6hring", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Torsten", | |
| "middle": [], | |
| "last": "Marek", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Yvonne", | |
| "middle": [], | |
| "last": "Samuelsson", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2010, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Martin Volk, Anne G\u00f6hring, Torsten Marek, and Yvonne Samuelsson. 2010. SMULTRON (version 3.0) -The Stockholm MULtilingual parallel TRee- bank. electronic.", | |
| "links": null | |
| }, | |
| "BIBREF25": { | |
| "ref_id": "b25", | |
| "title": "Grammatik der deutschen Sprache. Schriften des Instituts f\u00fcr deutsche Sprache", | |
| "authors": [ | |
| { | |
| "first": "Gisela", | |
| "middle": [], | |
| "last": "Zifonun", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Ludger", | |
| "middle": [], | |
| "last": "Hoffmann", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Bruno", | |
| "middle": [], | |
| "last": "Strecker", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 1997, | |
| "venue": "", | |
| "volume": "7", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Gisela Zifonun, Ludger Hoffmann, and Bruno Strecker. 1997. Grammatik der deutschen Sprache. Schriften des Instituts f\u00fcr deutsche Sprache; 7.1-3. de Gruyter, Berlin; New York.", | |
| "links": null | |
| } | |
| }, | |
| "ref_entries": { | |
| "TABREF1": { | |
| "text": "", | |
| "html": null, | |
| "content": "<table><tr><td colspan=\"6\">: Distribution of semantic functions of auf</td></tr><tr><td colspan=\"6\">(on) in relation to the syntactic function. The syn-</td></tr><tr><td colspan=\"6\">tactic function \"predicative\" is labelled as \"p\".</td></tr><tr><td colspan=\"5\">sem\\syn vmod mod opp ?mod</td><td>-</td></tr><tr><td>verbal</td><td>16</td><td colspan=\"2\">8 107</td><td>2</td><td>1</td></tr><tr><td>nominal</td><td>4</td><td>53</td><td>7</td><td/><td/></tr><tr><td>coll</td><td>3</td><td/><td>1</td><td/><td/></tr><tr><td>TEM</td><td>4</td><td/><td/><td>1</td><td/></tr><tr><td>MOD</td><td>46</td><td>2</td><td>2</td><td>6</td><td>4</td></tr><tr><td>INS</td><td>75</td><td>3</td><td>4</td><td>1</td><td>1</td></tr><tr><td>ORN</td><td>5</td><td>56</td><td/><td>4</td><td>1</td></tr><tr><td>COM</td><td>30</td><td>10</td><td>3</td><td>2</td><td>1</td></tr><tr><td>IDE</td><td>8</td><td>1</td><td/><td>7</td><td/></tr><tr><td>SIZ</td><td>4</td><td>6</td><td>1</td><td>1</td><td/></tr><tr><td>?</td><td>1</td><td>3</td><td/><td>1</td><td>3</td></tr><tr><td/><td>196</td><td colspan=\"2\">142 125</td><td colspan=\"2\">25 11</td></tr></table>", | |
| "num": null, | |
| "type_str": "table" | |
| }, | |
| "TABREF2": { | |
| "text": "", | |
| "html": null, | |
| "content": "<table><tr><td>: Distribution of the semantic function of</td></tr><tr><td>mit in relation to the syntactic function. The syn-</td></tr><tr><td>tactic function predicative is not shown in the table</td></tr><tr><td>because it appeared only once.</td></tr><tr><td>were available, we could probably map their</td></tr><tr><td>scheme to our scheme. No attempt was made by</td></tr><tr><td>M\u00fcller et al. (2011) to learn a model for preposi-</td></tr><tr><td>tion classification based on their semantic classes.</td></tr><tr><td>Their approach based on logistic regression as de-</td></tr><tr><td>scribed in Kiss et al. (2010) focuses on determiner</td></tr><tr><td>omission in PPs.</td></tr><tr><td>The work of</td></tr></table>", | |
| "num": null, | |
| "type_str": "table" | |
| }, | |
| "TABREF4": { | |
| "text": "Translations (Google Translate) of German mit in English and Spanish. Columns and rows are ordered by margin frequencies.", | |
| "html": null, | |
| "content": "<table><tr><td>es\\en</td><td>on</td><td>to</td><td>0</td><td>in</td><td colspan=\"5\">at of for about by with around</td></tr><tr><td>en</td><td>182</td><td>7</td><td colspan=\"3\">7 27 17</td><td/><td>1</td><td/><td>1</td><td>2 244</td></tr><tr><td>a</td><td>8</td><td>71</td><td>6</td><td colspan=\"2\">2 10</td><td>2</td><td>2</td><td/><td>1</td><td>102</td></tr><tr><td>0</td><td>7</td><td colspan=\"2\">7 33</td><td/><td/><td>3</td><td>6</td><td>1</td><td>57</td></tr><tr><td>de</td><td>9</td><td>10</td><td>2</td><td>5</td><td colspan=\"2\">1 17</td><td>3</td><td/><td>47</td></tr><tr><td>sobre</td><td>15</td><td>2</td><td/><td/><td/><td/><td/><td>1</td><td>18</td></tr><tr><td>para</td><td/><td>8</td><td/><td/><td/><td/><td>4</td><td/><td>12</td></tr><tr><td>por</td><td>5</td><td/><td>1</td><td/><td/><td/><td>1</td><td/><td>1</td><td>8</td></tr><tr><td>con</td><td>1</td><td>2</td><td/><td/><td>1</td><td/><td/><td/><td>1</td><td>5</td></tr><tr><td>contra</td><td>2</td><td/><td/><td/><td/><td/><td/><td/><td>2</td></tr><tr><td/><td colspan=\"6\">229 107 49 34 29 22</td><td>17</td><td>2</td><td>2</td><td>2</td><td>2 495</td></tr></table>", | |
| "num": null, | |
| "type_str": "table" | |
| }, | |
| "TABREF5": { | |
| "text": "Translations of German auf in English and Spanish. Translations appearing only once are not shown.", | |
| "html": null, | |
| "content": "<table/>", | |
| "num": null, | |
| "type_str": "table" | |
| }, | |
| "TABREF9": { | |
| "text": "Performance of feature sets for semantic classification accuracy: auf (N = 171). The following classes are considered: LOC, DIR, MOD, TLOC, CAU, TEM.", | |
| "html": null, | |
| "content": "<table><tr><td>Evidence</td><td>Mean</td><td colspan=\"2\">SD \u2206absbs</td><td>\u2206relbs</td></tr><tr><td>baseline</td><td>26.2</td><td>9.9</td><td/><td/></tr><tr><td>head</td><td>27.6</td><td>8.8</td><td>+1.4</td><td>+5.3</td></tr><tr><td>context</td><td colspan=\"2\">36.6 12.3</td><td>+10.3</td><td>+39.5</td></tr><tr><td>neighbor</td><td colspan=\"2\">39.3 13.7</td><td>+13.1</td><td>+50.0</td></tr><tr><td>syntax</td><td>42.1</td><td>4.5</td><td>+15.9</td><td>+60.5</td></tr><tr><td>head/neigh./en/es</td><td colspan=\"2\">40.7 11.3</td><td>+14.5</td><td>+55.3</td></tr><tr><td>head/syntax/neigh.</td><td>52.4</td><td>5.7</td><td colspan=\"2\">+26.2 +100.0</td></tr></table>", | |
| "num": null, | |
| "type_str": "table" | |
| }, | |
| "TABREF10": { | |
| "text": "Performance of feature sets for semantic classification accuracy: mit (N = 290). The following classes are considered: TEM, MOD, INS, ORN, COM, IDE, SIZ.", | |
| "html": null, | |
| "content": "<table/>", | |
| "num": null, | |
| "type_str": "table" | |
| } | |
| } | |
| } | |
| } |