| { | |
| "paper_id": "1993", | |
| "header": { | |
| "generated_with": "S2ORC 1.0.0", | |
| "date_generated": "2023-01-19T07:34:19.690333Z" | |
| }, | |
| "title": "The Interplay of Syntactic and Semantic Node Labels in Partial Parsing", | |
| "authors": [ | |
| { | |
| "first": "David", | |
| "middle": [ | |
| "D" | |
| ], | |
| "last": "Mcdonald", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "", | |
| "institution": "Brandeis University", | |
| "location": { | |
| "addrLine": "14 Brantwood Road", | |
| "postCode": "02174", | |
| "settlement": "Arlington", | |
| "region": "MA", | |
| "country": "USA" | |
| } | |
| }, | |
| "email": "" | |
| } | |
| ], | |
| "year": "", | |
| "venue": null, | |
| "identifiers": {}, | |
| "abstract": "Our natural language comprehension system, \"Sparser\" , uses a semantic grammar in conjunc tion with a domain model that defines the categories and already-known individuals that can be expected in the sublanguages we are studying, the most significant of which to date has been articles from the Wall Street Journal's \"Who's News\" column. In this paper we describe the systematic use of default syntactic rules in this grammar: an alternative set of labels on consitu tents that are used to capture generalities in the semantic interpretation of constructions like the verbal auxiliaries or many adverbials. Syntactic rules form the basis of a set of schemas in a Tree Adjoining Grammar that are used as templates from which to create the primary, semantically labeled rules of the grammar as part of defining the categories in the domain models. This design permits the semantic grammar to be developed on a linguistically principled basis since all the rules must conform to syntactically sound patterns.", | |
| "pdf_parse": { | |
| "paper_id": "1993", | |
| "_pdf_hash": "", | |
| "abstract": [ | |
| { | |
| "text": "Our natural language comprehension system, \"Sparser\" , uses a semantic grammar in conjunc tion with a domain model that defines the categories and already-known individuals that can be expected in the sublanguages we are studying, the most significant of which to date has been articles from the Wall Street Journal's \"Who's News\" column. In this paper we describe the systematic use of default syntactic rules in this grammar: an alternative set of labels on consitu tents that are used to capture generalities in the semantic interpretation of constructions like the verbal auxiliaries or many adverbials. Syntactic rules form the basis of a set of schemas in a Tree Adjoining Grammar that are used as templates from which to create the primary, semantically labeled rules of the grammar as part of defining the categories in the domain models. This design permits the semantic grammar to be developed on a linguistically principled basis since all the rules must conform to syntactically sound patterns.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Abstract", | |
| "sec_num": null | |
| } | |
| ], | |
| "body_text": [ | |
| { | |
| "text": "The rate-limiting step in advancing how much machines can understand from a natural language text is the size and thoroughness of the seman tic models used to make sense of the information the texts contain. At the same time, the state of the art in parsing and information extraction mandates that the texts being analyzed can no longer be artificial -made up by linguists to test their grammatical theories -but should be actual texts written by people for other people; indeed, the right kind of test corpus today will be downloaded from an online service on the day of the test. To accomodate this combination of unre stricted sources and limited semantic compe tence, one focuses on sublanguage analysis (in the sense of Kittredge & Lehrberger 1982) . This may be simply a restriction in the source of one's texts to just one topic and register, such as in the present instance single paragraph articles taken from the \"Who's News\" column of the Wall Street Journal, which contain almost nothing beyond announcements of the promotions or retirements of executives in commercial businesses. It may more broadly be a concentration on a given kind of information and its standard sublanguage but with no constraint on the sources in which it appears, e.g. extracting executive job change information in articles of any length appearing anywhere in the Journal or other business news sources available online, while at the same time functioning robustly dispite the unknown words and unparseable constructions in other portions of the texts.", | |
| "cite_spans": [ | |
| { | |
| "start": 726, | |
| "end": 754, | |
| "text": "Kittredge & Lehrberger 1982)", | |
| "ref_id": null | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Partial parsing and se mantic grammars", | |
| "sec_num": "1" | |
| }, | |
| { | |
| "text": "This then is the sence in which the \"par tial parsing\" of this paper's title should be un derstood: the parses the system makes, and its comprehension of the parsed segments, should be both complete and thorough, but those seg ments will typically constitute only a portion of the whole text being analyzed. Since the intru sion of off-topic segments can be at any granu larity -apposatives or adjuncts within clauses, modifiers within NPs, etc., a partial parser in this sense will incorporate some heuristic techniques for compensating for such gaps in its analysis, b u t for the most part its parsing machinery can be conventional and its only substantial differences from a parser intended to always parse every part of its text will lie in its grammar.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Partial parsing and se mantic grammars", | |
| "sec_num": "1" | |
| }, | |
| { | |
| "text": "In order to constrain the grammar to only those topics and sublanguages within the corpus that the system's semantic models will under stand, one is naturally drawn to the use of se mantic grammars. Here the grammar will only provide an analysis if the model will comprehend it -no parse node will be allowed to be formed if it does not also have a semantic interpretation. No effort is wasted in trying to make sense of portions of the text not on the target topic since the grammar will leave them unanalyzed and no interpretations will be initiated.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Partial parsing and se mantic grammars", | |
| "sec_num": "1" | |
| }, | |
| { | |
| "text": "Introduced originally by Dick Burton as the basis of the natural language interface to the So phie ICAI system (Burton 1976; Burton & Brown 1975) and later used for the interfaces in Lad der (Hendrix et al. 1978) , Planes (Waltz et al. 1976) , and many other systems, semantic gram mars have the advantage of high specificity, easy, almost common-sense definition, and extremely low run-time ambiguity. At the same time, they also suffer from brittleness, minimal carryover and significant effort in extending them to other domains, and a striking lack of ability to appre ciate linguistic generalizations.", | |
| "cite_spans": [ | |
| { | |
| "start": 111, | |
| "end": 124, | |
| "text": "(Burton 1976;", | |
| "ref_id": null | |
| }, | |
| { | |
| "start": 125, | |
| "end": 145, | |
| "text": "Burton & Brown 1975)", | |
| "ref_id": null | |
| }, | |
| { | |
| "start": 191, | |
| "end": 212, | |
| "text": "(Hendrix et al. 1978)", | |
| "ref_id": null | |
| }, | |
| { | |
| "start": 215, | |
| "end": 241, | |
| "text": "Planes (Waltz et al. 1976)", | |
| "ref_id": null | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Partial parsing and se mantic grammars", | |
| "sec_num": "1" | |
| }, | |
| { | |
| "text": "In this paper we will describe a version of se mantic grammar that addresses these problems while retaining the formalism's strengths. We provide a systematic basis for relating the seman tic categories used by the parser to the categories and representational structures used in the se mantic model that provides the text's interpre tation. We also provide the parser with a paral lel set of standard syntactic categories shadowing the semantic categories and providing the basis of domain-independent default rules with a common semantics. This is all brought together in a set of general schemas, derived from a Tree Adjoining Grammar, that allow the definition of the gram mar to b\ufffd done as part of the definition of the se mantic model, ensuring consistency and reducing much of the labor of writing semantic grammars in the process.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Partial parsing and se mantic grammars", | |
| "sec_num": "1" | |
| }, | |
| { | |
| "text": "We will begin by motivating the way in which", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Partial parsing and se mantic grammars", | |
| "sec_num": "1" | |
| }, | |
| { | |
| "text": "we link parsing rules to their interpretations, and with that introduce the notion of schemas that are used to define the two simultaneously. We will then describe the 'form' rules that provide the syntactic defaults, and present the algorithm for checking for possible constituent combinations in some detail since it is the means by which the two kinds of rules are interleaved.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "McDONALD", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "What is the goal of a parser? The customary an swer is that it is to determine the form of a textits decomposition into constituents and the gram matical relations among them, which is to then serve as the basis of its semantic interpretation -the recovery and representation of its mean ing. The line between form and interpretation begins to blur, however, once one adopts a seman tic grammar as the basis of the analysis of text form. A constituent in a semantic grammar has a label that reflects its category when taken as an thing or event-type in the world: 'person' rather than NP, 'elect' or 'get-position' rather than verb. Given a multi-level analysis by the parser, with transformations or their equivalent to render a surface-level representation into a canonical form, such a choice of categorizations will yield a repre sentation not too disimilar from some conceptions of logical form, with the constituent labels pro viding the predicate names applied to variables introduced by the transformations. The present work, however, employs a denota tional rather than descriptive notion of semantic interpretation. Here the process of parsing a text incrementally accesses or contructs a set of unique semantic entities -individuals in a model of the world and the ongoing discourse as represented in the mind (code) of the language user. The inter pretation function is applied rule by rule as the parse proceeds, rather than having to wait for it to end. Many of the individuals and all of the base types of this model will exist before the text is processed, and the task of the language under standing system is taken to be to recover from its analysis of the text the objects representing these individuals, and to add to the model any newly established individuals and any new relations or properties over them and pre-existing individuals.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Parsing to objects", | |
| "sec_num": "2" | |
| }, | |
| { | |
| "text": "As a parser then, the Sparser system does re-cover a phrase structure-based analysis of a text's form . But this form is only a means to an end, namely to project from the words of the text, as channeled by the constituent structure, directly to the objects in the world/discourse model with out any interveening intermediate level of repre sentation of the meaning as another expression such as a logical form. Since this projection hap pens incrementally as the parse proceeds from be ginning to end in one pass through the text, the parser may as well be said to be parsing to objects (in something of the manner of Reis beck and Mar tin 1985), since it will not complete a constituent without also recovering that constituent's coun terpart in the model. This being the case, and given the goal of a partial parser to tightly coordinate what the parser's grammar is able to analyze with what its semantic model is able to understand, a natural way to proceed is to have the rules of the gram mar written as a side-effect of defining the types of the model. What this means is that when one defines, say, a category of event like 'confirm-in posi tion', that definition should include the def inition of the rules needed to parse texts about that event-type. (Confirm-in-position is intended here as the denotation of the verb \"confirm\" as used in a sentence like \"J. Gordon Strasser, act ing president and chief executive officer of this gold mining company, was confirmed in the posts ... \", which is typical of the \"Who's News\" texts we have worked with.) Concombinantly, the pars ing rules created by this definition should spell out not only how the constituents of the text com pose, but also how those compositions should be projected to the discourse model to incrementally construct an individual to represent the confirma tion event rule by rule as the parse progresses.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Parsing to objects", | |
| "sec_num": "2" | |
| }, | |
| { | |
| "text": "Let us consider what has to go into these def initions so that establishing 'confirm-in-position' as an event-type in the model will at the same time provide the rules that the grammar needs to handle a sentence like the example. For its own purposes, the semantic definition of an event or other relations will specify a set of necessary and optional participants, including value restrictions constraining what categories of individuals those participants can be. For 'confirm-in-position', as evidenced by news articles in the Wall Street Journal, the agent that causes a confirmation can be either a company or its board of directors; the participant who gets confirmed is a person; and that person is confirmed to some position. We can use these categories directly to provide the labels for the corresponding constituents in the semantic grammar, where they take on the force of selectional restrictions.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Parsing to objects", | |
| "sec_num": "2" | |
| }, | |
| { | |
| "text": "Beyond the selectional restrictions on the par ticipants, the definitions must supply the words that have the intended meaning as one of their senses (i.e. the semantic object under definition) as well as the syntactic su bcategorization frames that the words take. Given the choices of lexical realization, syntactic information is then needed to spell out the phrasal patterns corresponding to the subcategorizations and how the participants project onto them.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Parsing to objects", | |
| "sec_num": "2" | |
| }, | |
| { | |
| "text": "With few exceptions, the subcategorization frames are shared by a great many words, and consequently can be schematized and shared among the definitions ( e.g. transitive verb, in transitive, noun taking quantification of number, noun holding a position in a sequence, etc.). The patterns of verbs especially are most naturally given in terms of syntactic phrasal categories, and can thus also provide the default syntactic cat egories that are provided with the rules of the semantic grammar, as we shall see.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Parsing to objects", | |
| "sec_num": "2" | |
| }, | |
| { | |
| "text": "We fill the requirement to fit all of the partic ipants of a relation to a single sharable syntac tic schema by adopting Lexicalized Tree Adjoin ing Grammars as the linguistic formalism (Joshi 1985; Shabes, Abeille & Joshi 1988) . TA G's ex tended domain of locality allows the configura tional and grammatical relations of all of a lexi cal head's arguments to be given at once, which is important here since the semantic relationship of each argument to its matrix relation must be specified in the schema, and this is easier to do if all of the participants of a domain relation can be paired with their syntactic counterparts in one statement. The schemas we adopt are a nota tional variant on the TAG Tree Families of Abeille 1988, which give all of the transformational vari ations of a subcategorization frame in one struc ture. The particulars are described in section five.", | |
| "cite_spans": [ | |
| { | |
| "start": 186, | |
| "end": 198, | |
| "text": "(Joshi 1985;", | |
| "ref_id": null | |
| }, | |
| { | |
| "start": 199, | |
| "end": 228, | |
| "text": "Shabes, Abeille & Joshi 1988)", | |
| "ref_id": null | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Parsing to objects", | |
| "sec_num": "2" | |
| }, | |
| { | |
| "text": "To summarize the framework so far, we are employing in the Sparser system a grammar of rules of syntactic form which employ semantic categories as their primary labels, with standard syntactic labels provided with them as defaults. Each rule has an accompanying semantic inter- pretation function that will find or construct the denotation of the constituent it creates. The grammar rules and the set of model-level object types are created at the same time by definitions that recruit the value restrictions on the partic ipants in a relation to supply the semantic la bels on the grammar's rules, where they implicitly serve the role of selectional restrictions ensuring that the rules cannot complete unless their con stituents will make sense to the world model. The set of rules created by a type definition is deter mined by instantiating the TAG tree family(s) corresponding to the subcategorization frames of the word or words that can realize the relation type in a text.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Parsing to objects", | |
| "sec_num": "2" | |
| }, | |
| { | |
| "text": "The effect on the final grammar is as though we had taken a conventional syntactic grammar that used syntactic preterminal categories and performed two multiplications: the first takes its preterminal categories for content words and mul tiplies them across all of the lexical choices spec ified by the definitions of the model's domain specific types. The second takes the rules com posing normal syntactic nonterminals and multi plies them across all of the sets of selectional re strictions dictated by the semantic categories of the individuals that can participate in the domain relations. The resulting grammar is dramatically larger than the original, but it will show a equally dramatic increase in runtime efficiency because of its minimal ambiguity and its immediate and in cremental connection to the semantic interpreta-. tion, which, after all, is the ultimate purpose of the grammar in the first place.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Parsing to objects", | |
| "sec_num": "2" | |
| }, | |
| { | |
| "text": "Before introducing the notation and mechanisms of the model-level definitions and the subcatego rization schemas, we should look at some exam ples of the grammar rules -productions -that Sparser uses. We will start with some semanti cally labeled rules based on a type definitionthe output of a schema, and then look at the syn tactically labeled rules that complement it . We start with the set of rules produced for 'confirm in-position', all of which are listed just in terms of their semantic labels in figure 1 We will then look at the details of what is represented in one of these rules and show how the syntactic and denotational information is represented and put to use. We will end the section by sketching how the rules are used in a parse of a text that is com pletely within the job-change sublanguage so that one can get a sense of the interplay between form and denotation in the Sparser system.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Integrating rewrite rules and denotations", | |
| "sec_num": "3" | |
| }, | |
| { | |
| "text": "#<phrase-structure-rule psr11 :lefthand-side #<category conf irm-in-position> :righthand-side ( #<word ''confirmed' '> ) :syntactic-form #<category verb+ed> :referent #<category confirm-in-position> > Figure 2 : Rewrite rule 11 #<edge e22 ;; a designator to use as the edge 's name :starts-at #<position 18> :ends-at #<posit ion 19> :category #<category confirm-in-position> :left-daughter #<word ''confirmed' '> :right-daughter :unary-rule ,, i.e. only the 'left ' daughter matters :used-in nil ;; later points to the edge this is a daughter of :form #<category verb+ed> :referent . #<category confirm-in-position> > Figure 1 ) consists of five unary lexical rules (9 -13), and eight binary rules over non-terminals corresponding to the conventional rules: S -> NP VP, VP -> VG NP, S -> NP VG, S -> S PP, and two NP rules for nominaliza tions. These semantically labeled rules are com plemented by a set of generically applicable syn tactically labeled 'form rules', e.g. for the \"be + ed\" of the passive, other auxiliaries and standard adverbs, and for the definite or indefi nite articles that can occur with the nominalizations. We take these up in a moment. 1", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 201, | |
| "end": 209, | |
| "text": "Figure 2", | |
| "ref_id": null | |
| }, | |
| { | |
| "start": 618, | |
| "end": 626, | |
| "text": "Figure 1", | |
| "ref_id": "FIGREF0" | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Integrating rewrite rules and denotations", | |
| "sec_num": "3" | |
| }, | |
| { | |
| "text": "In parsing our example, \"J. Gordon Strasser, acting president and chief executive offi cer of this gold mining company, was confirmed in the posts\" , the instantiation of a model-level 'confirm-in-position' event will start when the word \"confirmed\" is scanned, firing rewrite rule 11 (figure 2). That rule has the internal repre sentation shown below. The '#<>' notation indi cates pointers to objects, with the first term in the brackets indicating the object's type and the rest a summarizing print form, typically the sym bol or string used to name it.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Integrating rewrite rules and denotations", | |
| "sec_num": "3" | |
| }, | |
| { | |
| "text": "Note that the 'referent' given with the rule is a pointer directly into the system's semantic model. It is available as part of the statement of the rule because the rule is created as a side-effect of defining that event-type. Notice also that the rule also indicates the default syntactic label ( cat egory) that is to be given to the constituents it forms, i.e. verb+ed, complementing the primary semantic label given as the rule's lefthand side.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Integrating rewrite rules and denotations", | |
| "sec_num": "3" | |
| }, | |
| { | |
| "text": "When completed, this rule creates a nonter minal node -an edge in Sparser's chart -as shown in figure 3. For purposes of illustration we have the example text start at chart position zero, which puts the word \"confirmed\" between posi tions 18 and 19. (In the actual text, this was the first clause of the first sentence of a one paragraph article, with an extensive set of headers and with the name and location of the company where the change was going on announced in a dateline just 1 There are not as many phrasal rules in this set as one might expect for a transitive verb given the notion of a transformation family. This is because Sparser handles the relative pronouns of relative clauses by rewriting them with the semantic label of their NP head, making them equivalent here to rules 2 and 5, and it treats other WH constructions using 'form rules' as described below. Other predictable constructions such as it-clefts or topicalization have not yet occured anywhere in the studied portions of the Who's News corpus and have been omitted just because we have not had the examples to study for their semantic analysis.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Integrating rewrite rules and denotations", | |
| "sec_num": "3" | |
| }, | |
| { | |
| "text": "before the sentence started -all of which was also entered into the chart and processed.)", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Integrating rewrite rules and denotations", | |
| "sec_num": "3" | |
| }, | |
| { | |
| "text": "Note that the edge essentially just carries over the information given in the rule. The category given as the 'lefthand-side' of the rule becomes the edge's 'category' -it is this semantic label which will be the primary label used the course of parsing. The category given as the 'syntactic form' of the rule becomes the edge's 'form' -its syntactic label, which will be used only if the se mantic label cannot participate in a composition with any of the neighbor edges, i.e. it is the de fault.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Integrating rewrite rules and denotations", | |
| "sec_num": "3" | |
| }, | |
| { | |
| "text": "The semantic analysis at work here holds that content words (nouns, verbs, adjectives) name categories of individuals in the semantic model. The denotation of \"confirmed\" is, as shown, the category ascribed to events of that type (which we also use as the semantic label on its edge, a com mon but not universal practice in this grammar since particularities of subcategorization and ap plicable substitution contexts can override it and call for more specific labelings). The denotation of the word \"company\" in the example is the cat egory used for all individuals representing compa nies and similar enterprises; \"posts\" denotes the category for positions, specialized in this case to note that more than one title and/or company is involved; and so on.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Integrating rewrite rules and denotations", | |
| "sec_num": "3" | |
| }, | |
| { | |
| "text": "The composition of modifiers onto a head con stituent may specialize the category of the head to one of its subcategories (as done implicitly with the plurality of \"posts\"), or may predicate at tributes to an individual. The composition of some discourse anchor -tense for verbs or de-terminers for nouns -will change the denotation to now pick out some individual fitting the cate gory of the phrase accumulated on the headline to that point. This happens when \"was\" and \"con firm\" are composed, where we will establish an individual of domain-type (category) 'confirm-in position'. This involves a 'rule of form' which we take up in the next section.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Integrating rewrite rules and denotations", | |
| "sec_num": "3" | |
| }, | |
| { | |
| "text": "In the interests of space we will jump ahead to the point when the major constituents of the example have been formed so that we cari con centrate on how the rules for confirm-in-position are applied and how they take the new individual in the model -an instance of this event-typeand fill its participant roles with the individuals denoted by those constituents.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Integrating rewrite rules and denotations", | |
| "sec_num": "3" | |
| }, | |
| { | |
| "text": "Below (figure 4) is a vertical presentation of that chart state, followed by a list of the edges with their semantic and syntactic labels and their denotations. A compact notation is used here for illustration purposes. The edges names, e.g., 'e23', reflect the order in which the edges were formed. The numbers around the excerpted texts are the chart positions.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Integrating rewrite rules and denotations", | |
| "sec_num": "3" | |
| }, | |
| { | |
| "text": "At this point SPARSER will apply rule 5 from the set above to combine edges e20 and e23 to form a clause ('S') labeled 'job-event' from posi tion 1 to position 20, edge e30. When that edge is formed, the denotations of its daughter edges are combined following the interpretation func tion given with the rule (see section 5). The re sult, the denotation of e30, will be the same in dividual denoted by e23, but now that individual is augmented by having its 'agent ' role bound to the person pointed to by the subject edge e20. This new edge, e30, will then be combined with the pp-adjunct edge e29 using rule 3 with a similar result. Another edge, e31, is formed as shown in figure 5. Its referent is again the same individual that was instantiated when the verb group edge was formed (i.e. the individual repre senting this instance of confirm-in-position). Now it gets its 'position' role bound to the individual of type position denoted by e29.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Integrating rewrite rules and denotations", | |
| "sec_num": "3" | |
| }, | |
| { | |
| "text": "The parse is now finished syntactically since we have recovered a grammatically complete unit, a sentence. Semantically it is still incomplete, since the text did not explicitly give the agent that did the confirming and so the event-type is not yet saturated. In this genre it would be safe to infer that the agent is the company where the person now holds the position.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Integrating rewrite rules and denotations", | |
| "sec_num": "3" | |
| }, | |
| { | |
| "text": "The preponderance of the syntactic ( surface form) rewrite rules in Sparser's grammar will have words or semantic categories on their righthand sides resulting from the application of some TAG schema during the definition of a domain-specific type in the world model; when the distinction must be made we will refer to them as 'semanti cally characterized rules'. There are also rewrite rules whose categories are taken from the conven tional syntactic vocabulary and consequently are shared across all of the semantic domains; we call these 'form rules' because they make reference to general aspects of surface form rather ideosyn cratic aspects that apply to only certain domains of discourse and their sublanguages. Fo rm rules are the subject of this section. We will look at an example of a form rule and the rationale behind its use, and then go through the 'Check' algo rithm that coordinates the selection of semanti cally characterized or form rules.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Form rules", | |
| "sec_num": "4" | |
| }, | |
| { | |
| "text": "A form rule combines a word or occasionally a semantically labeled edge -call it the 'literal' -with the form category of another edge -call that edge the 'head'. The head edge will always lie on the headline of the final tree; the literal will always be a specifier, an auxiliary, or a modify ing adjunct or adverb. The purpose of a form rule is to capture a general fact about phrase formation in the language, one whose implications for the semantics of the resulting phrase -its denota tion -is the same regardless of the domain. \u2022", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Form rules", | |
| "sec_num": "4" | |
| }, | |
| { | |
| "text": "The implementation of form rules in Sparser allows them to be transparent to the rest of the rules in the grammar. The new edge created by composing the literal and the head edge will have the same category ( or some linguistically principled projection of it) as that of its head edge -any semantically characterized rules that might eventually apply to the head edge will ap ply transparently to the new edge as though the #<form-rule fpsr1", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Form rules", | |
| "sec_num": "4" | |
| }, | |
| { | |
| "text": ":righthand-side ( #<word ''was' '> #<category verb+ed> ) :new-category :passive :syntactic-form #<category verb-group> :referent (:head :right-edge :subtype #<past> ) > Figure 6 : The applicable form rule for \"was confirmed\" #<edge e21 :starts-at #<posit ion 17> :ends-at #<position 18> :category #<word ''was''> :left-daughter #<word ''was''> :right-daughter :literal-in-a-rule :used-in nil :form #<category verbal-auxiliary> :referent nil > Figure 7 : The edge that spans the word \"was\" literal had not been there. The observent reader will already have noticed that the set of rules above was missing the cases needed to\" cover many likely texts: \" ... will be confirmed Wednesday in Seattle\", \". . . has been confirmed\", etc. All such combinations -with time and place adjuncts, with complex verbal auxiliaries -are handled uniformly throughout the grammar by form rules.", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 169, | |
| "end": 177, | |
| "text": "Figure 6", | |
| "ref_id": null | |
| }, | |
| { | |
| "start": 443, | |
| "end": 451, | |
| "text": "Figure 7", | |
| "ref_id": null | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Form rules", | |
| "sec_num": "4" | |
| }, | |
| { | |
| "text": "The exceptions to this uniformity come when there needs to be a domain-specific, semantic ( de notational) consequence to a particular combina tion, as when a verb subcategorizes specifically for a location. In such cases the grammar will simply include a semantically characterized rule for that combination, and the parser will look for that rule in preference to the form rule, as de scribed in section 4.1.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Form rules", | |
| "sec_num": "4" | |
| }, | |
| { | |
| "text": "The form rule that applies in the case of \"was confirmed\" is shown in figure 6. ", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 70, | |
| "end": 79, | |
| "text": "figure 6.", | |
| "ref_id": null | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Form rules", | |
| "sec_num": "4" | |
| }, | |
| { | |
| "text": "Before we can go into the formal algorithm for checking whether edges combine and coordi nating the two kinds of rules, we must also show the edge that spans the word \"was\" and discuss its possibly unusual aspects ( figure 7) .", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 216, | |
| "end": 225, | |
| "text": "figure 7)", | |
| "ref_id": null | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "(Note that the referent in this rule is not a direct pointer into the model as before but a schematic func tion. It will take the edge on the right of the two constituents -edge-22 denoting the category 'confirm-in-position -and look in the model for one of its specializations (one of its subtypes) that represents the event having occurred in the past, which will then become the denotation of the new edge.) Informally speaking, this rule applies in the present example because the word \"was\" is ad jacent to edge-22 which has the form category 'verb+ed', and because there is no semantically characterized rule that combines \"was\" and edge-22's primary category, 'confirm-in-position '. The new edge formed by the completion of the rule will have the label 'confirm-in-position+passive', as dictated by the form rule's 'new-category ' field in coordination with the schema that de fined confirm-in-position and made provisions for that projection from the original category over the verb. (Had there been no new-category field in the rule, then by default the new edge would have taken the same label as the one on the head edge -edge-22.)", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "Note that the label on this edge (its 'cate gory' field) contains a word rather than a cate gory. There are two reasons for this. The first is that allowing words as nonterminals permits an elegant treatment of abbreviations: One can simply rewrite the combination of the abbrevi ation and its period as the full, unabbreviated word; thereby avoiding any need to redundantly extend grammar to include combination rules for the abbreviation as well as for the full word ( e.g. \"Company\" -> \"Co\" \".\"). If the same abbrevi ation has several expansions then multiple edges will be put into the chart.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "(Note that the referent in this rule is not a direct pointer into the model as before but a schematic func tion. It will take the edge on the right of the two constituents -edge-22 denoting the category 'confirm-in-position -and look in the model for one of its specializations (one of its subtypes) that represents the event having occurred in the past, which will then become the denotation of the new edge.) Informally speaking, this rule applies in the present example because the word \"was\" is ad jacent to edge-22 which has the form category 'verb+ed', and because there is no semantically characterized rule that combines \"was\" and edge-22's primary category, 'confirm-in-position '. The new edge formed by the completion of the rule will have the label 'confirm-in-position+passive', as dictated by the form rule's 'new-category ' field in coordination with the schema that de fined confirm-in-position and made provisions for that projection from the original category over the verb. (Had there been no new-category field in the rule, then by default the new edge would have taken the same label as the one on the head edge -edge-22.)", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "The second reason is to simplify the implemen tation of the parsing algorithm so that it needs to manipulate only a single data type -edgeswhen looking for rule applications, rather than both edges and words. Any word that is men tioned in a rewrite rule will be covered with an edge when it is scanned, the label on that edge being the word itself. This also permits a simple way to discriminate between known and unknown words in the course of a parse. A known word ( one with rules in the grammar) will always have one or more edges over it, if only a trivial edge labeled with that word; unknown words will not have edges, leaving a gap among the preterminal edges of the chart, which can be used to trigger heuristics for dealing with unknown words when they appear in certain contexts.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "(Note that the referent in this rule is not a direct pointer into the model as before but a schematic func tion. It will take the edge on the right of the two constituents -edge-22 denoting the category 'confirm-in-position -and look in the model for one of its specializations (one of its subtypes) that represents the event having occurred in the past, which will then become the denotation of the new edge.) Informally speaking, this rule applies in the present example because the word \"was\" is ad jacent to edge-22 which has the form category 'verb+ed', and because there is no semantically characterized rule that combines \"was\" and edge-22's primary category, 'confirm-in-position '. The new edge formed by the completion of the rule will have the label 'confirm-in-position+passive', as dictated by the form rule's 'new-category ' field in coordination with the schema that de fined confirm-in-position and made provisions for that projection from the original category over the verb. (Had there been no new-category field in the rule, then by default the new edge would have taken the same label as the one on the head edge -edge-22.)", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "We should also note that all of Sparser's ac tions are taken over unary\" or binary rules. Rules with more than two terms on their righthand sides are converted to a set of rules in Chomsky Nor mal Form, using a dotted-rule convention to pro vide names for the intermediate non-terminal la bels created in the process. Also, in describing the Check algorithm we will not go into any de tail about the larger aspects of the parsing algo rithm that determines when or whether two ad jacent edges will be checked. Basically the parser operates bottom up and left to right, forming constituents from their heads outwards; however, overlaying that conventional traversal of the space of possible constituents is a moderately complex control structure that reduces the search to a de terministic algorithm, in the sense that every edge formed will be part of the final analysis and every edge will have a unique parent edge in the tree. For some of the particulars see McDonald 1992.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "(Note that the referent in this rule is not a direct pointer into the model as before but a schematic func tion. It will take the edge on the right of the two constituents -edge-22 denoting the category 'confirm-in-position -and look in the model for one of its specializations (one of its subtypes) that represents the event having occurred in the past, which will then become the denotation of the new edge.) Informally speaking, this rule applies in the present example because the word \"was\" is ad jacent to edge-22 which has the form category 'verb+ed', and because there is no semantically characterized rule that combines \"was\" and edge-22's primary category, 'confirm-in-position '. The new edge formed by the completion of the rule will have the label 'confirm-in-position+passive', as dictated by the form rule's 'new-category ' field in coordination with the schema that de fined confirm-in-position and made provisions for that projection from the original category over the verb. (Had there been no new-category field in the rule, then by default the new edge would have taken the same label as the one on the head edge -edge-22.)", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "We can define Sparser's Check operation -its algorithm for seeing whether there is a rule in the grammar that combines two particular adjacent edges -as follows: Consider two adjacent edges which we will call \"Left\" and \"Right\" . Left ends at some position pi n the chart, and Right begins at this same po sition p. Left is said to be earlier in the chartcloser to the beginning of the text -because it begins at some position p' (p' being strictly less that p) and because Right ends at some position p\" (p\" being strictly greater than p ). Positions are indexed numerically starting with zero before the first word of the text and increasing in a pos itive direction, one position between each word. The position at which an edge begins always has a lower numerical index than the position at which it ends; no edge can begin and end at the same position.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "The algorithm that coordinates checking for both kinds of rules", | |
| "sec_num": "4.1" | |
| }, | |
| { | |
| "text": "Only adj acent edges can be combined to form a new edge. The new edge will start at position p' and end at p\" . The possibility of two adj acent edges combining is fu nction of their labels, where a label is a word or category in the 'category' field of an edge, or one of a specially designated set of 'form categories' pointed to from the 'form' field of an edge. The form field may be empty.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "The algorithm that coordinates checking for both kinds of rules", | |
| "sec_num": "4.1" | |
| }, | |
| { | |
| "text": "piled from the labels of the rules of the grammar. All rules are binary. The table will associate the righthand sides of those rules, taken as ordered pairs of labels, with the rules themselves. We use the table at runtime by forming an index from the labels of the adj acent constituents and look ing the index up in the table to see if that pair of constituents was one of the ones defined by the grammar. If it is, we return the rule located at that index in the table; if it is not, there will be no rule at that index and we return nil indicat ing that the edges do not combine. We presently implement the table as a hash on the numerical index; it would lend itself to a hardware imple mentation as will be clear from its definition.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "The Check operation is based on a table com", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "We form the table as follows. For each rule consider the two labels on its righthand side; call them LL and RL for the first and second (left and right) labels respectively. The rules are taken one at a time in some order: the textual order of their definition in the files of the grammar suffices. We assign each label two unique numbers. One num ber stands for the label when it appears as the left label of a binary rule (LL); the second stands for when it appears as the right label of a rule (RL fields of the two edges). For Left , retrieve the right-looking number that was assigned to itthe one it was given for when it appeared as the left of the two labels in a binary rule. Do the op posite for Right. Now add the two numbers, and use the sum as an index into the table. A rule will be returned if the two edges compose, and this rule is used as the basis of the new edge spanning them. If that index into the table is nil, indicating that no semantically characterised rule applies to that edge pair, then go on to check for combi nations involving form rules. In the present ex ample there will be no semantically characterised rule for \"was\" plus 'confirm-in-position', so we will move on to check form rules. There will always be two possible form rules to check for: one combining the category field la bel of Left and the form field label of Right; the other combining the form label of Left and the category label of Right. Only one of these two pairs will have an index defined for it because of the requirement on form rules that one of the two labels on their righthand sides be either a literal word or a semantic category. We start by com puting the table index for the first case, and only if that lookup into the table fails to return a rule do we try the second. If one or the other label fails to have a form option in its stored pair of index numbers then that case is undefined. If neither succeeds, then the two adjacent edges do not compose.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "The Check operation is based on a table com", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "This completes the description of Sparser's Check algorithm. It can be summarized as fol lows: Its purpose is to determine whether two adjacent edges can be composed to form a new edge. This is allowed in any of three cases: first if the semantic categories of the two edges match the righthand side of some rule in the grammar; failing that if either the semantic category of the first and the form category of the second corre spond to some form rule in the grammar, or visa versa. The test is made by adding numbers that were assigned to the labels as the grammar rules were defined, and using that index as the key into a table that records which pairs correspond to rules.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "The Check operation is based on a table com", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "The prohibition on having a form rule based on a pair of form labels is there to rule out the possibility of ever writing a conventional syntac tic grammar in this formalism. We rule this out for two reasons. The first is that allowing \"pure\" form rules would re-introduce to the grammar the massive ambiguities that syntactic grammars are prone to. The second is that doing so would miss the point of this grammar design, namely that the purpose of a grammar is to facilitate the map ping between form and meaning -between the edges formed over words or phrases and their de notations in the system's world model. We only allow form rules to be used if we are already con structing the projection to some individual in the model, i.e. if the combination is between an edge on the head .line which will have a projection and one of its adjuncts (broadly construed).", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "The Check operation is based on a table com", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "We have presented the Check algorithm in this much detail because we have seldom seen the rule lookup algorithms of other parsers described and feel that this kind of information should be added to the literature. In the present case, the very high efficiency of this constant-time lookup oper ation is especially important because of the very large number of rewrite rules in the grammar. A lookup algorithm that was sensitive to the size of the grammar ( or for that matter the presence of any 'order of the size of the grammar' operation within the parser) would strongly penalize the use a s.emantic grammar. Optimizing algorithms for speed is important even with today's processors because of the ever growing volume of text that must be processed. 2", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "The Check operation is based on a table com", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "As already discussed, the Sparser language com prehension system works in close consort with a universe of individuals and categories ( the indi vidualsU types) that supply the denotations of the words and phrases it parses: This model of the system's world (the topic domains it is compe tent in) and of the ongoing discourse is based on a set of type definitions, and we are using those def initions as the source of the semantically labeled rewrite rules that Sparser uses. In this section we will look at one such definition and the TAG schema that it uses to create the rules. We begin with a discussion of the schem\ufffd, shown in figure 8 on page 182. The 'binding-parameters' and 'labels' fields hold sets of symbols that will be substituted for when the schema is applied to the type-definition: The labels are used to define the rewrite patterns of the individual rewrite rules given in the 'cases' field; the parameters are placeholders for the def inition's participant roles. The initial terms in each case (e.g. :subject) are just indicators used by the grammar writer to help organize the cases.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Defining parsing rules as a side-effect of defining a do main object type", | |
| "sec_num": "5" | |
| }, | |
| { | |
| "text": "The parenthesized expressions following the indicators are the schematic form of the rules; left and righthand sides are indicated in the obvious way. The :head and : binds indicators after the syntactic part of each rule give its semantic in terpretation function. They indicate which of the two edges is the head and what role is filled by the denotation of the other edge.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Defining parsing rules as a side-effect of defining a do main object type", | |
| "sec_num": "5" | |
| }, | |
| { | |
| "text": "The label symbols on the lefthand sides of the rules ( e.g. S, VP, VG) designate the form labels that are to accompany the corresponding seman tic labels that are substitued for them to con struct the rules. An example of the final form of one of these rules was given earlier. job-event, e.g. elect, appoint, resign, etc The 'realization' field is the basis for instan tiating the category's portion of the semantic grammar. The rules constructed from it will rec ognize phrases denoting instances of individuals with this category in a text, and will guide the semantic interpretation to instantiate them and populate their participant roles. The field indi cates which syntactic rule schema should be used in this case the tree-family 'transitive/passive: shown just above. In its 'mapping' field it in dicates the correspondences between the roles of the category and the substitution variables in the schema. The 'main-verb' field (vs. 'head-noun' in other cases) indicates what word should be used as the lexical head of the tree family; its mor phological variants are calculated automatically, though if it were irregular the special cases would be given explicitly here. The 'saturation' field in dicates \u2022 that those three roles must be given in a text before an individual of this type is fully defined.", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 278, | |
| "end": 321, | |
| "text": "job-event, e.g. elect, appoint, resign, etc", | |
| "ref_id": null | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Defining parsing rules as a side-effect of defining a do main object type", | |
| "sec_num": "5" | |
| }, | |
| { | |
| "text": "When this form is executed to define the cat egory, its realization field is interpreted and the syntactic schema ( tree family) it indicates is ap plied to the parameters given in the field. Each of the rule cases of the schema is taken up in turn and the corresponding rewrite rule created by making the substitutions called for by the mapping field. Notice that there are two addi tional cases given in this category's realization field. They identify how the category's 'position' role is filled, namely with either of two different adjunct prepositional phrases, one using \"in\" , the other \"as\" . (They are analyzed here as attaching to the sentence rather than the verb phrase for compatibility with the way nominalizations are treated.) This kind of local augmentation of the general syntactic schemas ( the set of tree fami lies) is convenient for specifying the often ideosyn cratic way that individual verbs can subcategorize for optional complements.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "The first several fields of confirm-in-position are what you would expect in a modern frame lan guage. The category 's position in the taxonomic specialization hierarchy is given by its 'special izes' field. Its 'binds' field lists the participants that are specific to this event-type; although an individual instance can have additional partici pants like time \u2022 and place by warrant of inherit ing from the category TeventU. The participants' value-restrictions are indicated by the :v /r flag. The 'instantiates' field indicates which category above it in the hierarchy an instance should be indexed under in the discourse history; thus for purposes of searching the discourse history, a con firm event is taken to be equivalent to any other", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "Let us look at one case, the first one in the schema that specifies S -> NP-subject VP. (The NP is annotated as the subject to distinguish it among the substitution variables from the other NP, the direct object.) Looking to tlie mapping specification we see we have to make two rules because the subject np can take either of two values. We form the lefthand side of the both by substituting the category 'job-event' for S. We form the righthand side of one by substituting the category 'board-of-directors' for the variable 'np subject' and substituting 'job-event\\agent' for 'vp'; the other rule gets 'company' substituted for its np-subject.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "The first several fields of confirm-in-position are what you would expect in a modern frame lan guage. The category 's position in the taxonomic specialization hierarchy is given by its 'special izes' field. Its 'binds' field lists the participants that are specific to this event-type; although an individual instance can have additional partici pants like time \u2022 and place by warrant of inherit ing from the category TeventU. The participants' value-restrictions are indicated by the :v /r flag. The 'instantiates' field indicates which category above it in the hierarchy an instance should be indexed under in the discourse history; thus for purposes of searching the discourse history, a con firm event is taken to be equivalent to any other", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "The rule must also have a syntactic-form cat egory, which will be the form label of the edge created when the rule completes. It takes this from the syntactic category (substitution vari able) given as the rule's lefthand side in the schema (i.e. S). The rule .talso must have a ref erent, which it gets from the corresponding part of the case in the schema, subsituting its role 'per son' for the symbol 'patient' given in the schema, again as indicated by t;h\ufffd mapping field.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "The first several fields of confirm-in-position are what you would expect in a modern frame lan guage. The category 's position in the taxonomic specialization hierarchy is given by its 'special izes' field. Its 'binds' field lists the participants that are specific to this event-type; although an individual instance can have additional partici pants like time \u2022 and place by warrant of inherit ing from the category TeventU. The participants' value-restrictions are indicated by the :v /r flag. The 'instantiates' field indicates which category above it in the hierarchy an instance should be indexed under in the discourse history; thus for purposes of searching the discourse history, a con firm event is taken to be equivalent to any other", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "We should note in\u2022\u2022 closing this section that a number of largely arbitrary decisions were made in the writing of this grammar, which is to say in the selection of what semantic categories to use in stating the correspondences between semantic labels and syntactic labels in the mapping field of a definition like this. For example we have made a generalization about the semantic interpretations of a whole semantic class of verbs by deciding to specify the VP label as 'job-event\\agent' rather than, say, 'confirm-in-position' (which would be as specific as we would be able to be). By choos ing the more general category (which should be read as in a categorial grammar: look to compose with an agent to the left of this constituent and label the new edge 'job-event'), we are saying that all the verbs in this family ( or rather their model level denotations) make the same interpretation of their subject NP, i.e. they all map it to their equivalent of the agent role. This points out that linguistic generalizations can be made in the se mantic realm as well as the syntactic, and that by using a semantic grammar we are able to express this in a very direct fashion.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": ".\u2022 \u2022 ,", | |
| "sec_num": "1" | |
| }, | |
| { | |
| "text": "Two substantive complaints have been made against semantic grammars in the literature ( e.g. . Wallace 1984). One is that they cannot be easily extended as one moves from one domain to the next, while a thorough syntactic grammar can be taken over unchanged. This observation misses a crucial point, however, namely that to under stand a text one must find the correspondence between the structural analysis provided by the syntax and the individuals and domain-types in the world model of the application system that is going to use the information given in the text. There is no question that the world model of the first domain that one works with will have to be extended when one moves to the next -the next domain will involve new kinds of objects and new types of relations, and these will have to be added to the model if texts in the new domain are to be understood. There is no other possibility and the fact that there may be no change in the syntactic grammar is besides the point because it is not the whole story. By using a semantic grammar and having it constructed directly from the world model as done here with Sparser, we will able to keep the grammar in step with the extensions to the model. In particular, we will be continually re minded, as we work on the model, that we need to work out the words and syntactic construc tions used to talk about the new concepts, since each domain type should have a realization field to indicate this information.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Concluding remarks", | |
| "sec_num": "6" | |
| }, | |
| { | |
| "text": "The second substantive complaint is that se mantic grammars are incapable of capturing lin guistic generalizations, and so one cannot take advantage of these uniformities when writing the grammar, and one may be tempted to write rules that would be unreasonable from a linguistic per spective and so likely to be brittle in practice ( e.g. defining constituents that combine an np with a following preposition, leaving the prepo sition's complement stranded). This is the more legitimate of the two complaints, but it is also specifically what the present system was designed to address.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Concluding remarks", | |
| "sec_num": "6" | |
| }, | |
| { | |
| "text": "By keeping the grammarian ( domain modeler) from writing rules directly except in the most ideosyncratic cases ( e.g. for dates written as \"3/31/93\" ), and instead forcing them to work by way of some well-crafted syntactic schemas, we have imposed a discipline on the grammar, forc ing it to fit our conception of properly analyzed linguistic structure as captured in the schemas. At the same time we have reduced the overall ef fort required, since it is markedly more dispatch ful to copy and specialize the realization fields of a set of related domain-type definitions than to write out the rules individually.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Concluding remarks", | |
| "sec_num": "6" | |
| }, | |
| { | |
| "text": "We have been using semantic grammars in the Sparser system since its inception three years ago. Form rules were introduced after the experiences with its first major grammar, and had a signif icant effect in simplifying the effort to develop new vocabulary since they allowed whole sets of rules for generic constructions to be used just by giving new semantic categories in the grammar the appropriate syntactic labels. The use of re-alization fields on model-level definitions and the development of the 'exploded tree families' used as schemas is comparatively recent, and we have only now redone the original grammar for the Who's News domain in its terms. The proof of the pudding, as it were, will come as the models and grammars for the next topic domains are added in the coming months (these will be for joint ventures and quarterly earnings). If the effort to make these extensions is dramatically smaller than when the original grammars were developed by hand we will judge the design a pragmatic suc cess. If a large family of syntactic schemas, espe cially for noun phrases, can be developed that addresses the bulk of the construction-types that we see in the business texts we work with then we will have gotten a long way towards solving the problem of extending grammars to new domains.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Concluding remarks", | |
| "sec_num": "6" | |
| }, | |
| { | |
| "text": "One day of the Wall Street Journal averages about half a megabyte. It is available online at about six o'clock in the morning. We will want the results of our information extractions and their pragmatic analysis on the desks of clients when they come in three hours later, and it is only one of any number of online text-based information services that might be used. Sparser processes the sublanguage-rich portions of a text at about 20 words a second(Mac-II, 68020, 15.7mhz); off-topic portions run at about 100 words/sec.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "", | |
| "sec_num": null | |
| } | |
| ], | |
| "back_matter": [ | |
| { | |
| "text": "A", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Abeille, A. (1988)", | |
| "sec_num": null | |
| } | |
| ], | |
| "bib_entries": {}, | |
| "ref_entries": { | |
| "FIGREF0": { | |
| "text": "Semantically labeled rules", | |
| "type_str": "figure", | |
| "num": null, | |
| "uris": null | |
| }, | |
| "FIGREF1": { | |
| "text": "An edge in the chart This rule set (", | |
| "type_str": "figure", | |
| "num": null, | |
| "uris": null | |
| }, | |
| "FIGREF2": { | |
| "text": "confirm-in\ufffdposition+passive , vg , #<individual : confirm-in-position> ] [e29 : in-position, pp , #<individual :position : collect ive President I> ] Chart state and list of edges", | |
| "type_str": "figure", | |
| "num": null, | |
| "uris": null | |
| }, | |
| "FIGREF3": { | |
| "text": "TAG schema for rule creation", | |
| "type_str": "figure", | |
| "num": null, | |
| "uris": null | |
| }, | |
| "FIGREF4": { | |
| "text": "The domain-specific model-level type 'confirm-in-position'", | |
| "type_str": "figure", | |
| "num": null, | |
| "uris": null | |
| } | |
| } | |
| } | |
| } |