Datasets:

WINGNUS
/

ACL-OCL

	{
	"paper_id": "1988",
	"header": {
	"generated_with": "S2ORC 1.0.0",
	"date_generated": "2023-01-19T07:44:31.189779Z"
	},
	"title": "Lexical Realization in Natural Language Generation",
	"authors": [
	{
	"first": "Sergei",
	"middle": [],
	"last": "Nirenburg",
	"suffix": "",
	"affiliation": {
	"laboratory": "",
	"institution": "Carnegie Mellon University",
	"location": {}
	},
	"email": ""
	},
	{
	"first": "Rita",
	"middle": [],
	"last": "Mccardell",
	"suffix": "",
	"affiliation": {
	"laboratory": "",
	"institution": "Carnegie Mellon University",
	"location": {}
	},
	"email": ""
	},
	{
	"first": "Eric",
	"middle": [],
	"last": "Nyberg",
	"suffix": "",
	"affiliation": {
	"laboratory": "",
	"institution": "Carnegie Mellon University",
	"location": {}
	},
	"email": ""
	},
	{
	"first": "Scott",
	"middle": [],
	"last": "Huffman",
	"suffix": "",
	"affiliation": {
	"laboratory": "",
	"institution": "Carnegie Mellon University",
	"location": {}
	},
	"email": ""
	},
	{
	"first": "Edward",
	"middle": [],
	"last": "Kenschaft",
	"suffix": "",
	"affiliation": {
	"laboratory": "",
	"institution": "Carnegie Mellon University",
	"location": {}
	},
	"email": ""
	},
	{
	"first": "Irene",
	"middle": [],
	"last": "Nirenburg",
	"suffix": "",
	"affiliation": {},
	"email": ""
	}
	],
	"year": "",
	"venue": null,
	"identifiers": {},
	"abstract": "This paper describes a procedure for lexical selection of open-class lexical items in a natural language generation system. An optimum lexical selection module must be able to make realization decisions under varying contextual circumstances. First, it must be able to operate without the influence of context, based on meaning correspondences between elements of conceptual input and the lexical inventory of the target language. Second, it must be able to use contextual constraints, as supported by collocational information in the generation lexicon. Third, there must be an option of realizing input representations pronominally or through definite descriptions. Finally, there must also be an option of using elliptical constructions. The nature of background knowledge and the algorithm we suggest for this task are described. The lexical selection procedure is a part of a comprehensive generation system, DIOGENES.",
	"pdf_parse": {
	"paper_id": "1988",
	"_pdf_hash": "",
	"abstract": [
	{
	"text": "This paper describes a procedure for lexical selection of open-class lexical items in a natural language generation system. An optimum lexical selection module must be able to make realization decisions under varying contextual circumstances. First, it must be able to operate without the influence of context, based on meaning correspondences between elements of conceptual input and the lexical inventory of the target language. Second, it must be able to use contextual constraints, as supported by collocational information in the generation lexicon. Third, there must be an option of realizing input representations pronominally or through definite descriptions. Finally, there must also be an option of using elliptical constructions. The nature of background knowledge and the algorithm we suggest for this task are described. The lexical selection procedure is a part of a comprehensive generation system, DIOGENES.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Abstract",
	"sec_num": null
	}
	],
	"body_text": [
	{
	"text": "1 Our place on the generation research map.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "",
	"sec_num": null
	},
	{
	"text": "Natural language generation is traditionally divided into two stages: the utterance planning ('what to say') stage and the lexical and syntactic realization ('how to say it') stage. The latter stage consists, essentially, of a large set of realization choices for the various meanings of the input, using the morphological, syntactic and lexical means of expression in the target language (TL). Research reported here deals with the process of lexical selection during this second stage of generation. Many of the existing generation systems have been conceived as components of natural language interfaces to database systems. In such generators the lexical inventory can be strongly constrained without jeopardizing the quality of the interaction (cf, e.g., McKeown, 1985) . Such systems necessarily concentrate on choosing appropriate TL syntax -indeed, generators are expected to produce adequate syntactic structures. Lexical selection becomes more important when it is difficult to constrain the types of output in generation, and, consequently, when the lexicon becomes large. Machine translation and automatic text summarization are among applications that by nature require a wide range of outputs and have to use a sizable lexicon. Note that of these two, the former does not involve utterance planning and concentrates on lexical and syntactic realization.",
	"cite_spans": [
	{
	"start": 749,
	"end": 774,
	"text": "(cf, e.g., McKeown, 1985)",
	"ref_id": null
	}
	],
	"ref_spans": [],
	"eq_spans": [],
	"section": "",
	"sec_num": null
	},
	{
	"text": "In the natural language generation community, attention to the task of lexical selection attention has recently grown. Of course, it has always been recognized as an important problem (cf. Danlos, 1984; Jacobs, 1985; Bienkowski, 1986 ; and the survey Cumming, 1986) and was addressed in a well-known early generation project (Goldman, 1975 ). However, most major NLG efforts of the early 1980s (e.g., Mann and Matthiessen, 1985; McDonald, 1983; McKeown, 1985 , Appelt, 1985 concentrated more on syntactic and stylistic realization as well as to text planning than to lexical selection. Among those who dealt with lexical selection was the SEMSYN project (R\u00f6sner, 1986) . Currently, work on lexical selection is going on at the University of Pennsylvania (e.g., Marcus, 1987) and at University of California at Berkeley (Ward, 1988) . Still, the set of problems facing this field is significant. One motivation for our research was that we agree with Marcus (1987, p. 211 ) that at present 'most generation systems don't use words at all,' and we believe that the quality of generation output will improve significantly once an adequate lexical selection component becomes a standard part of a NLG system.",
	"cite_spans": [
	{
	"start": 189,
	"end": 202,
	"text": "Danlos, 1984;",
	"ref_id": "BIBREF4"
	},
	{
	"start": 203,
	"end": 216,
	"text": "Jacobs, 1985;",
	"ref_id": "BIBREF9"
	},
	{
	"start": 217,
	"end": 233,
	"text": "Bienkowski, 1986",
	"ref_id": "BIBREF1"
	},
	{
	"start": 251,
	"end": 265,
	"text": "Cumming, 1986)",
	"ref_id": "BIBREF3"
	},
	{
	"start": 325,
	"end": 339,
	"text": "(Goldman, 1975",
	"ref_id": "BIBREF6"
	},
	{
	"start": 401,
	"end": 428,
	"text": "Mann and Matthiessen, 1985;",
	"ref_id": "BIBREF13"
	},
	{
	"start": 429,
	"end": 444,
	"text": "McDonald, 1983;",
	"ref_id": "BIBREF15"
	},
	{
	"start": 445,
	"end": 458,
	"text": "McKeown, 1985",
	"ref_id": "BIBREF15"
	},
	{
	"start": 459,
	"end": 473,
	"text": ", Appelt, 1985",
	"ref_id": "BIBREF0"
	},
	{
	"start": 654,
	"end": 668,
	"text": "(R\u00f6sner, 1986)",
	"ref_id": "BIBREF22"
	},
	{
	"start": 761,
	"end": 774,
	"text": "Marcus, 1987)",
	"ref_id": "BIBREF14"
	},
	{
	"start": 819,
	"end": 831,
	"text": "(Ward, 1988)",
	"ref_id": null
	},
	{
	"start": 950,
	"end": 970,
	"text": "Marcus (1987, p. 211",
	"ref_id": null
	}
	],
	"ref_spans": [],
	"eq_spans": [],
	"section": "",
	"sec_num": null
	},
	{
	"text": "Research reported in this paper was performed within the DIOGENES project (Nirenburg, 1987) , whose objective is to provide a high-quality generator for a knowledge-based interlingual machine translation system. The input to this generator is a set of a) world concept instances that represent the propositional content of the original text, and b) sets of text parameter values that represent its pragmatic content (see Section 4 for more details). In this paper we deal with a subset of the generation task, namely, the selection of open-class lexical items to realize the meanings of object, event and property tokens in the input. Thus, the output of the generation module described here is a lexical unit or a pronoun in the target language.",
	"cite_spans": [
	{
	"start": 74,
	"end": 91,
	"text": "(Nirenburg, 1987)",
	"ref_id": "BIBREF19"
	}
	],
	"ref_spans": [],
	"eq_spans": [],
	"section": "The Task",
	"sec_num": "2"
	},
	{
	"text": "Our approach (and especially the expected input) to text generation is similar to that of the SEMSYN project (e.g., R\u00f6sner, 1986) . Lexical selection is not, however, an immediate concern of and is not discussed at any length in SEMSYN descriptions (see Laubsch et al., 1984, p. 492) , and a published analysis of practical difficulties encountered by the project (Hanakata et al., 1986) does not address this issue at all. Furthermore, since until very recently that project had to generate sentence-length texts (article titles), the problem of definite descriptions, pronominalization and ellipsis did not become acutely important.",
	"cite_spans": [
	{
	"start": 116,
	"end": 129,
	"text": "R\u00f6sner, 1986)",
	"ref_id": "BIBREF22"
	},
	{
	"start": 254,
	"end": 283,
	"text": "Laubsch et al., 1984, p. 492)",
	"ref_id": null
	},
	{
	"start": 364,
	"end": 387,
	"text": "(Hanakata et al., 1986)",
	"ref_id": "BIBREF7"
	}
	],
	"ref_spans": [],
	"eq_spans": [],
	"section": "The Task",
	"sec_num": "2"
	},
	{
	"text": "Lexical choice is not a straightforward task. Suppose we have to express in English the meaning 'a person whose sex is male and whose age is between 13 and 15 years.' What knowledge do people use in order to come up with an appropriate choice out of such candidate realizations as those listed in (1).",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Why is it a difficult task?",
	"sec_num": "3"
	},
	{
	"text": "(1) boy, kid, teenager, youth, child, young man, schoolboy, adolescent, man Without a sentential context the choice, based on closeness of the meaning match and generality of meaning, should be boy. For a computer program to be capable of making choices like this, it has to possess a preference-assigning capability on the matches between the meanings of the candidate lexical realization on the one hand and the input meaning unit (see the discussion of the matching metric below) on the other.",
	"cite_spans": [
	{
	"start": 9,
	"end": 75,
	"text": "kid, teenager, youth, child, young man, schoolboy, adolescent, man",
	"ref_id": null
	}
	],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Why is it a difficult task?",
	"sec_num": "3"
	},
	{
	"text": "Lexical choices are, however, typically made in context. Contextual relations among lexical units reflect meaning-induced constraints on cooccurrence (selectional restrictions: admire takes a human subject). Sometimes, however, it is difficult to formulate a cooccurrence constraint in terms of selectional restrictions alone. Thus, for example, the causative construction with the English influence requires exert; its Russian equivalent vlijanie requires okazyvat', and the latter is not a Russian correlate of exert other than in the above and very few similar syntagmatic constructions. Why do we use, in English, shed with tears or leaves but don't usually say shed water out of a bucket or they drop tears every time when <...>? Such properties of the lexical stock of a natural language are called collocational. We will now illustrate the concept of collocation through several examples of 'semantic' or syntagmatic collocations.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Collocations",
	"sec_num": "3.1"
	},
	{
	"text": "Consider the conceptual operator of a large quantity of, a (relative) value for measuring quantities (of materials, forces, qualities, properties, etc.). It is realized in English in accordance with collocational properties of the lexical units that are used as its operands. Not every quantity goes with every realization of the above operator. Members of the set <big, enormous, great, high, large, strong, wide> of potential realizations of a large quantity of can cooccur with every one of the members of the set <amount, difficulty, expanse, selection, voltage> of quantities. We say high voltage but a large amount. It would be inappropriate for a generation system to produce something like high selection or large difficulty. (Note that in parsing the problem of assigning a similar semantic marker to all the various expressions from the example can, in principle, be tackled through a mechanism of metaphor processing, whereby a general heuristic rule is developed for processing metaphorical input belonging a single class, such as, for instance, a large quantity of... -see Lakoff and Johnson, 1980 , for an extensive listing of potential metaphor classes; in generation, however, the task is the opposite -to produce fluent metaphorical language. Since this depends not on regularities of meaning, but rather on the idiosyncrasies of meaning realization in the various natural languages, the general rules will be more difficult to come by and formulate.)",
	"cite_spans": [
	{
	"start": 1086,
	"end": 1110,
	"text": "Lakoff and Johnson, 1980",
	"ref_id": "BIBREF10"
	}
	],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Collocations",
	"sec_num": "3.1"
	},
	{
	"text": "An additional class of collocations are the paradigmatic collocations. These are best exemplified by the 'set-complement' collocations such as the English left and right or parents and children. The knowledge of these collocations, for instance, is necessary for the process of lexical selection of conjoined constructions, such as ladies and gentlemen.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Collocations",
	"sec_num": "3.1"
	},
	{
	"text": "Collocational relations are defined on lexical units, not meaning representations. The study of collocations ascends to Firth (1951) ; it is a central part of the Meaning -Text school of linguistics -cf. Mel'\u010duk, 1974; 1981 . The importance of collocational properties in generations has been recognized (cf. Cumming, 1986) , but relatively few systems actually include collocational information in their decision processes.",
	"cite_spans": [
	{
	"start": 120,
	"end": 132,
	"text": "Firth (1951)",
	"ref_id": "BIBREF6"
	},
	{
	"start": 204,
	"end": 218,
	"text": "Mel'\u010duk, 1974;",
	"ref_id": "BIBREF16"
	},
	{
	"start": 219,
	"end": 223,
	"text": "1981",
	"ref_id": "BIBREF17"
	},
	{
	"start": 309,
	"end": 323,
	"text": "Cumming, 1986)",
	"ref_id": "BIBREF3"
	}
	],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Collocations",
	"sec_num": "3.1"
	},
	{
	"text": "Certain contexts completely alleviate the problem of open-class lexical selection. Consider the following (gloss of an) input segment",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Ellipsis and Anaphora",
	"sec_num": "3.2"
	},
	{
	"text": "(2) Clause 1 : Buy(John 3 book 7 ), time 1 , focus: book 7",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Ellipsis and Anaphora",
	"sec_num": "3.2"
	},
	{
	"text": "Clause 2 : Bring(John 3 book 7 office 1 ), belong-to(office 1 John 3 ), time 2 : time 2 > time 1 , focus: office 1 Clause 3 : Read(John 3 book 7 ), aspect: inchoative, time 3 : after(time 2 )",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Ellipsis and Anaphora",
	"sec_num": "3.2"
	},
	{
	"text": "One of the adequate ways of realizing it is:",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Ellipsis and Anaphora",
	"sec_num": "3.2"
	},
	{
	"text": "(3) John bought a book. He brought this book to his office and started to read it.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Ellipsis and Anaphora",
	"sec_num": "3.2"
	},
	{
	"text": "There are seven instances of the three object-type concepts in the case-role slots of the input propositions above. Each of the three concepts is realized lexically only once. In two cases these meanings were realized through pronominalization and in one each through definite description and an elliptical construction. This example shows that non-lexical realization is an integral part of the process of lexical selection in generation.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Ellipsis and Anaphora",
	"sec_num": "3.2"
	},
	{
	"text": "In what follows we briefly describe the system architecture, the knowledge structures and the algorithm we use for selecting open-class lexical items in generation.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Ellipsis and Anaphora",
	"sec_num": "3.2"
	},
	{
	"text": "DIOGENES is a distributed natural language generation system featuring a blackboard-type control structure. The processing in it is concentrated in the knowledge sources which are triggered by the state of the various blackboards. The latter contain the input to generation as well as all intermediate and final results of DIOGENES operation, represented uniformly in the frame-oriented knowledge representation language FrameKit Joseph, 1985 and Nyberg, 1988) . The implementation vehicle for DIOGENES is CMU CommonLisp running on an IBM PC RT. (Description of the DIOGENES control structure and data handling is beyond the scope of this paper.) Background knowledge in DIOGENES includes the following components relevant to the task of lexical selection:",
	"cite_spans": [
	{
	"start": 430,
	"end": 446,
	"text": "Joseph, 1985 and",
	"ref_id": "BIBREF2"
	},
	{
	"start": 447,
	"end": 460,
	"text": "Nyberg, 1988)",
	"ref_id": null
	}
	],
	"ref_spans": [],
	"eq_spans": [],
	"section": "The System and the Knowledge",
	"sec_num": "4"
	},
	{
	"text": "\u2022 a concept lexicon, a set of knowledge structures that describe object and event-types in the (sub)world of the texts to be generated (the first application of DIOGENES is, for example, in the domain of computer hardware manuals)",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "The System and the Knowledge",
	"sec_num": "4"
	},
	{
	"text": "\u2022 a generation lexicon that links (sub)world concepts (or, more accurately, their instances) with particular lexical units of the target language.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "The System and the Knowledge",
	"sec_num": "4"
	},
	{
	"text": "The above description is necessarily incomplete. See Nirenburg, 1987 for an extensive specification of all the facets of DIOGENES.",
	"cite_spans": [
	{
	"start": 53,
	"end": 68,
	"text": "Nirenburg, 1987",
	"ref_id": "BIBREF19"
	}
	],
	"ref_spans": [],
	"eq_spans": [],
	"section": "The System and the Knowledge",
	"sec_num": "4"
	},
	{
	"text": "The input to DIOGENES, known as interlingua text (ILT), is a set of FrameKit frames representing the propositional and non-propositional meanings of the source language text input into the translation system (see Nirenburg et al., 1986 Nirenburg et al., , 1987b Nirenburg and Carbonell, 1987 for a detailed description). The ILT is produced during the analysis stage of the MT process by the parser, the semantic interpreter and, if needed, with human help. This latter stage is needed because our requirements for the input to generation are such that no current analysis system can produce them in a completely automatic fashion.",
	"cite_spans": [
	{
	"start": 213,
	"end": 235,
	"text": "Nirenburg et al., 1986",
	"ref_id": "BIBREF18"
	},
	{
	"start": 236,
	"end": 261,
	"text": "Nirenburg et al., , 1987b",
	"ref_id": "BIBREF20"
	},
	{
	"start": 262,
	"end": 291,
	"text": "Nirenburg and Carbonell, 1987",
	"ref_id": "BIBREF19"
	}
	],
	"ref_spans": [],
	"eq_spans": [],
	"section": "The System and the Knowledge",
	"sec_num": "4"
	},
	{
	"text": "The following is a sample InterLingua Text.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "The System and the Knowledge",
	"sec_num": "4"
	},
	{
	"text": "(make-frame clausel (clauseid (value clausel)) (propositionid (value propositionl)) (speechactid (value speech-actl))) (make-frame propositionl (propositionid (value propositionl)) (clauseid (value clausel)) (process-type (value action)) (is-token-of (value throw)) (aspectid (value aspectl)) (space (value propositionl.space)) (time (value (at propositionl.time))) (subworld (value everyday-world)) (modality (value real)) (manner (value quickly)) (agent (value rolel)) (object (value role2)) (instrument (value role3)) (source (value rolel)) (destination (value role4))) (make-frame aspectl (phase (value end)) (iteration (value 1) ) ;in seconds (duration (value 1))) (make-frame speech-actl (speech-act (value assertion)) (direct? (value no)) (speaker (value author)) (hearer (value reader)) (time (value speech-actl.time))) (make-frame rolel (roleid (value rolel)) (clauseid (value clausel)) (comment (value \"a young boy\")) (referent (value personl)) (description (value rolel)) (is-token-of (value person)) (age (value (10 11))) (sex (value male))) (make-frame role2 (roleid (value role2)) (clauseid (value clausel)) (comment (value \"a green rubber ball\")) (referent (value balll)) (description (value role2)) (is-token-of (value ball)) (color (value green)) (made-of (value rubber))) (make-frame role3 (roleid (value role3)) (clauseid (value clausel)) (comment (value arm)) (referent (value arml)) (description (value role3)) (is-token-of (value arm)) (part-of (value rolel))) (make-frame role4 (roleid (value role4)) (clauseid (value clausel)) (comment (value \"a big fast car\")) (referent (value vehiclel)) (description (value role4)) (is-token-of (value vehicle)) (medium (value road)) (propulsion (value fuel)) (wheels (value 3 4 ) ) (size (value big)) (velocity (value (50 1 8 0 ) ) ) )",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "The System and the Knowledge",
	"sec_num": "4"
	},
	{
	"text": "Given the above ILT, DIOGENES produces the text",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "The System and the Knowledge",
	"sec_num": "4"
	},
	{
	"text": "The young boy quickly threw a green rubber ball at a big fast car.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "The System and the Knowledge",
	"sec_num": "4"
	},
	{
	"text": "::= ( <meaning-pattern> <TL-pattern>* ) <meaning-pattern> ::= ((is-token-of (value <CL-concept>)) [(<relation> (value ",
	"cite_spans": [
	{
	"start": 98,
	"end": 117,
	"text": "[(<relation> (value",
	"ref_id": null
	}
	],
	"ref_spans": [],
	"eq_spans": [],
	"section": "GL-entry",
	"sec_num": null
	},
	{
	"text": "The main static knowledge source for the generating of open-class items is a specialized generation lexicon (GL). The structure of an entry in the generation lexicon in DIOGENES is shown in Figure 1 (the BNF is incomplete wherever obvious).",
	"cite_spans": [],
	"ref_spans": [
	{
	"start": 190,
	"end": 198,
	"text": "Figure 1",
	"ref_id": null
	}
	],
	"eq_spans": [],
	"section": "The Generation Lexicon",
	"sec_num": "4.1"
	},
	{
	"text": "The importance value serves to distinguish the saliency of the various relations for the identity of the entry head. Thus, for instance, generating youth instead of boy seems to be less a deviation than generating girl. This is why the importance of the sex slot in the example below is greater than that of the age slot.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "The Generation Lexicon",
	"sec_num": "4.1"
	},
	{
	"text": "(The sample GL entries below do not contain a full complement of collocation relations.) (make-frame boy (is-token-of (value person)) (sex (value male) (importance 10)) (age (value (2 15)) (importance 4)) (lexeme (value \"boy\")) (syntactic-info (lexical-class noun) (noun-type class)) (morphological-info (plural regular)) (para-collocation (synonym lad kid child) (antonym girl adult) (hypernym person)) (syn-collocations-in (value boy.syn))) (make-frame boy.syn (agent-of (value play throw run jump) (strength 0.5)) (place (value school playground ball field) (strength 0.5))) (make-frame toss (is-token-of (value throw)) (direction (value up) (importance 3)) (altitude (value (70 150))) ;in feet (importance 3)) (velocity (value (0 30))) ;in miles/hour (importance 9)) (lexeme (value \"toss\")) (syntactic-info (lexical-class verb) (verb-type transitive)) (morphological-info (past-tense regular) (past-participle regular)) (para-collocation (antonym catch) (synonym cast propel toss fling hurl pitch)) (syn-collocations-in (value toss.syn))) (make-frame toss.syn (agent (value gambler) (strength 0.7) ) (object (value coin) (strength 0 . 7 ) )) (make-frame fast (is-token-of (value velocity)) (velocity (percent-of-range (51 8 0 ) ) ) (lexeme (value \"fast\")) (syntactic-info (lexical-class adjective)) (morphological-info (comparative regular) (superlative regular)) (para-collocation (antonym slow) (synonym quick))) (make-frame car (is-token-of (value vehicle)) (medium (value road) (importance 8)) (propulsion (value fuel) (importance 10)) (wheels (value 3 4) (importance 9)) (lexeme (value \"car\")) (syntactic-info (lexical-class noun) (noun-type class)) (morphological-info (plural regular)) (para-collocation (synonym automobile auto)) (syn-collocations-in (value car.syn))) (make-frame car.syn (object-of (value dent) (strength 0.8)) (size (value big) (strength 0.8)) (velocity (value fast) (strength 0 . 8 ) ))",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "The Generation Lexicon",
	"sec_num": "4.1"
	},
	{
	"text": "The ILT frames and the meaning patterns of GL entries against which they are matched are collections of slots whose fillers are members of domains (or value sets) predefined for each slot. The slot fillers can be .symbols, e.g., (temperature (value tepid)) or numbers, e.g., (weight (value 141.5)).",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "The Matching Metric",
	"sec_num": "4.2."
	},
	{
	"text": "The domains (or value sets) from which slot filler values are taken can be unordered, ordered discretely (hereafter, ordered) or ordered continuously (hereafter, continuous). An example of an unordered set would be occupations, an ordered set months-of-the-year and a continuous set height. Note that only the slot fillers of a continuous set can contain numeric values.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "The Matching Metric",
	"sec_num": "4.2."
	},
	{
	"text": "The cardinality of slot fillers in meaning patterns can be single as in (sex (value female)), enumerated as in (subjects (value physics math history)) or ranges as in (age (value 0 100)).",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "The Matching Metric",
	"sec_num": "4.2."
	},
	{
	"text": "The matching metric (see Nirenburg et al., 1987a) is designed on the basis of the following heuristics:",
	"cite_spans": [
	{
	"start": 25,
	"end": 49,
	"text": "Nirenburg et al., 1987a)",
	"ref_id": "BIBREF19"
	}
	],
	"ref_spans": [],
	"eq_spans": [],
	"section": "The Matching Metric",
	"sec_num": "4.2."
	},
	{
	"text": "\u2022 If the types of slot fillers or the domains of the filler values in the ILT and GL meaning pattern differ, then the match is declared inadequate and a maximum distance is assigned.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "The Matching Metric",
	"sec_num": "4.2."
	},
	{
	"text": "\u2022 The quality of the match is proportional to the size of the intersection of the actual slot fillers in the ILT and GL frames, which is then modified by the size of the domain itself.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "The Matching Metric",
	"sec_num": "4.2."
	},
	{
	"text": "\u2022 Each slot in a GL meaning pattern is rated with respect to its importance for the meaning expressed by the frame. A mismatch on a less important slot will have a smaller influence on the overall score of the match.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "The Matching Metric",
	"sec_num": "4.2."
	},
	{
	"text": "\u2022 The quality of a match between two frames is a weighted sum of the quality of the matches of the individual frame slots.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "The Matching Metric",
	"sec_num": "4.2."
	},
	{
	"text": "For the matching program, it is assumed that the ILT frame contains all the slots found in the GL meaning pattern. If this is not the case, it is assumed that a special module in the generator (called the augmentor, see Nirenburg, 1987) will infer the necessary slots and fill them with default values that are stored in the IL ontological hierarchy, i.e., the Concept Lexicon. Conversely, the ILT frame can (and regularly will) contain more slots than the GL frame. This is due to the fact that an ILT frame corresponds not to a single lexical unit, but rather to an entire phrase (such as a noun phrase, complete with the various modifiers). Thus, lexical selection proceeds in two phases: first, the heads of the phrases (typically, nouns and verbs) are generated and then the modifiers (typically, adjectives and adverbs, but frequently such recursive elements as prepositional phrases and relative clauses) are selected.",
	"cite_spans": [
	{
	"start": 220,
	"end": 236,
	"text": "Nirenburg, 1987)",
	"ref_id": "BIBREF19"
	}
	],
	"ref_spans": [],
	"eq_spans": [],
	"section": "The Matching Metric",
	"sec_num": "4.2."
	},
	{
	"text": "The matching program is the actual computer implementation of the heuristics listed in the previous section. It matches an ILT frame's meaning pattern against the meaning patterns of GL entries to determine the appropriate lexical item for the frame. The components of the GL entry that the matching program uses are the meaning pattern section (which is a frame in the ILT format) and the TL string. A modified GL entry can be seen in Figure 2 .",
	"cite_spans": [],
	"ref_spans": [
	{
	"start": 436,
	"end": 444,
	"text": "Figure 2",
	"ref_id": "FIGREF1"
	}
	],
	"eq_spans": [],
	"section": "The Matcher",
	"sec_num": "4.3"
	},
	{
	"text": "The meaning pattern portion of a GL entry is an instance of a frame, containing the specific slot fillers that should be present for the lexical item to be appropriate.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "The Matcher",
	"sec_num": "4.3"
	},
	{
	"text": "Since the matching program performs flexible matching, it must decide (unless there is a perfect match) exactly how far the input ILT frame is from a perfect match with the GL pattern. After all GL entries for a given concept (or token) such as person have been examined in this fashion, the one with the best score is selected. All other mismatches are assigned penalties proportional to their \"imperfectness\".",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "The Matcher",
	"sec_num": "4.3"
	},
	{
	"text": "((is-token-of (value *person)) (sex (value female) (importance 10)) (age (value (0 12)) (importance 6)) (lexeme girl) ) In this section, the actual empirically derived scoring function is described, along with the algorithm that uses the scoring function to create an aggregate score for all the slots in the ILT input frame and the particular GL entry.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "The Matcher",
	"sec_num": "4.3"
	},
	{
	"text": "For each slot which is present in BOTH the ILT and the GL entry, the fillers are compared. If the fillers are equal, then the slots match perfectly, and no penalty is assigned. If the fillers do not match perfectly, a penalty is computed for that slot. The function used to calculate the penalty depends on the type of slot filler (e.g., SINGLE, ENUMERATED, or RANGE). Once the penalty is computed, it is weighted according to the importance of that slot to a successful match, and the weighted number is added to the cumulative penalty for the match thus far.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "The Matcher",
	"sec_num": "4.3"
	},
	{
	"text": "The scoring for the various combinations of domain sets (there are 3 sets) and the cardinality of slot filler types (there are 3 types) are discussed in the next five sections. Note, however, that out of a maximum of 9, there are only 7 meaningful combinations:",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "The Matcher",
	"sec_num": "4.3"
	},
	{
	"text": "\u2022 Unordered, Single ",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "The Matcher",
	"sec_num": "4.3"
	},
	{
	"text": "An ordered domain with single element fillers can be exemplified by a set of single temperature 1 values in Figure 3 .",
	"cite_spans": [],
	"ref_spans": [
	{
	"start": 108,
	"end": 116,
	"text": "Figure 3",
	"ref_id": null
	}
	],
	"eq_spans": [],
	"section": "Ordered Domains, Single Element Fillers",
	"sec_num": "4.3.1"
	},
	{
	"text": "Suppose we are trying to match TEMPERATURE slots in an ILT frame and a GL entry:",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Ordered Domains, Single Element Fillers",
	"sec_num": "4.3.1"
	},
	{
	"text": "ILT: (temperature (value cool) ) GL: (temperature (value lukewarm))",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Ordered Domains, Single Element Fillers",
	"sec_num": "4.3.1"
	},
	{
	"text": "The penalty assigned to a mismatch depends on two variables:",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Ordered Domains, Single Element Fillers",
	"sec_num": "4.3.1"
	},
	{
	"text": "\u2022 D: the distance between the fillers in the ordered set of values (make-frame TEMPERATURE (instance-of (value slot)) (element-type (value symbol)) (domain-type (value ordered)) (cardinality (value single)) (elements (value (cold cool tepid lukewarm warm hot scalding)))) Figure 3 : The FrameKit Attribute Entry for TEMPERATURE",
	"cite_spans": [],
	"ref_spans": [
	{
	"start": 272,
	"end": 280,
	"text": "Figure 3",
	"ref_id": null
	}
	],
	"eq_spans": [],
	"section": "Ordered Domains, Single Element Fillers",
	"sec_num": "4.3.1"
	},
	{
	"text": "\u2022 C: the size of the domain of values",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Ordered Domains, Single Element Fillers",
	"sec_num": "4.3.1"
	},
	{
	"text": "The quality of the match depends, of course, on the distance between the two fillers; however, it is important to consider the size of the filler domain as well. The larger a domain, the higher the chance that the two fillers will diverge. Consequently, the penalty is lessened for larger domains, the distance between the fillers being equal. (If there are only two elements in the domain, there is a 50% chance of the correct filler appearing in the input, so a larger penalty is assigned if it does not.)",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Ordered Domains, Single Element Fillers",
	"sec_num": "4.3.1"
	},
	{
	"text": "A simple equation relates these two variables:",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Ordered Domains, Single Element Fillers",
	"sec_num": "4.3.1"
	},
	{
	"text": "where W is a numerical weight on the distance between the fillers and F is a damping function on the size of the domain. These parameters are added to the equation for calibration purposes (calibration is discussed in Section 4.3.7).",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Ordered Domains, Single Element Fillers",
	"sec_num": "4.3.1"
	},
	{
	"text": "The matching of single element fillers from unordered domains is quite similar to matching single filllers in ordered domains. The difference lies in the fact that the distance between the fillers is no longer meaningful, since they are not ordered in any way with respect to one another. The penalty assigned to the match becomes a function merely of the size of the domain (and hence the probability of the correct filler appearing):",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Unordered Domains, Single Element Fillers",
	"sec_num": "4.3.2"
	},
	{
	"text": "As stated before, only numbers can be represented in a continuous domain. The elements of the domain are defined by giving the endpoints of the domain (or closed interval) and the unit size of representation to be used in computing the distance between fillers. When defined in this manner, equation 1 can be used to compute the penalty for continuous domain sets as well. As an example, consider an alternate definition of the attribute TEMPERATURE in Figure 4 . As before, suppose we are trying to match temperature slots between an ILT and a GL frame:",
	"cite_spans": [],
	"ref_spans": [
	{
	"start": 453,
	"end": 461,
	"text": "Figure 4",
	"ref_id": null
	}
	],
	"eq_spans": [],
	"section": "Continuous Domains, Single Element Fillers",
	"sec_num": "4.3.3"
	},
	{
	"text": "ILT:(temperature (value 93.7)) GL: (temperature (value 98.6))",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Continuous Domains, Single Element Fillers",
	"sec_num": "4.3.3"
	},
	{
	"text": "Before evaluating equation 1, both the distance between the two fillers and the distance between the endpoints of the closed domain interval are computed using the defined unit-size value. Then the evaluation of equation 1 proceeds as before.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Continuous Domains, Single Element Fillers",
	"sec_num": "4.3.3"
	},
	{
	"text": "(make-frame TEMPERATURE (instance-of (value slot)) (element-type (value number)) (domain-type (value continuous)) (cardinality (value single)) (unit-size (value 1)) (elements (value (0 100)))) Figure 4 : A Continuous Attribute Entry for TEMPERATURE",
	"cite_spans": [],
	"ref_spans": [
	{
	"start": 193,
	"end": 201,
	"text": "Figure 4",
	"ref_id": null
	}
	],
	"eq_spans": [],
	"section": "Continuous Domains, Single Element Fillers",
	"sec_num": "4.3.3"
	},
	{
	"text": "Regardless of the domain, the function used to calculate the penalty for mismatched range fillers is identical to that for enumerated fillers. In the following discussion, it is assumed that the set operations of \"size of filler\" and \"size of filler intersection\" are executed directly on the enumerated fillers; the sizes and intersections of range fillers are calculated using the unit-size slot value. The abbreviation \"r/e\" is used to refer to the combination of both filler types, with the assumption that the appropriate preprocessing of ranges takes place prior to evaluation. When two r/e fillers do not match exactly, several factors are taken into account when calculating the size of the penalty: the size of each filler, the size of the intersection between the fillers, and the size of the domain.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Ordered Domains, Range/Enumerated Fillers",
	"sec_num": "4.3.4"
	},
	{
	"text": "The calculation of the filler size is straightforward. However, the size of the intersection between the fillers is much more relevant, and a little bit more difficult to obtain. The intersection is conceptually equivalent to D (the distance between the fillers in equation 1). However, the intersection can occur in one of four ways:",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Ordered Domains, Range/Enumerated Fillers",
	"sec_num": "4.3.4"
	},
	{
	"text": "1. The ILT filler and the GL filler overlap, and the resulting intersection is smaller than either filler.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Ordered Domains, Range/Enumerated Fillers",
	"sec_num": "4.3.4"
	},
	{
	"text": "2. The fillers overlap, and the ILT filler is contained within the GL filler.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Ordered Domains, Range/Enumerated Fillers",
	"sec_num": "4.3.4"
	},
	{
	"text": "3. The fillers overlap, and the GL filler is contained within the ILT filler.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Ordered Domains, Range/Enumerated Fillers",
	"sec_num": "4.3.4"
	},
	{
	"text": "Although the actual integer that represents the number of shared elements can be calculated by the same function for 1, 2, and 3 above, it is important for the matcher to distinguish between 2 and 3. In general, a match where the GL filler is contained within the ILT filler is better than a match where the ILT filler is contained within the GL filler. (On a certain abstract level, this is equivalent to a match where some slots are present in the GL but not in the ILT.)",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "The fillers do not overlap.",
	"sec_num": "4."
	},
	{
	"text": "When the fillers do not overlap, the intersection is indicated by a negative number that represents the distance between the nearest endpoints of the two fillers.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "The fillers do not overlap.",
	"sec_num": "4."
	},
	{
	"text": "The function used to calculate the penalties for ordered domains with r/e fillers employs the following variables:",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "The fillers do not overlap.",
	"sec_num": "4."
	},
	{
	"text": "\u2022 S i : the size of the ILT filler \u2022 S g : the size of the GL filler",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "The fillers do not overlap.",
	"sec_num": "4."
	},
	{
	"text": "\u2022 I: the size of the intersection between the two fillers where negative values indicate the distance between disjoint fillers The overall mismatch penalty (at the slot level) is computed by the following function:",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "The fillers do not overlap.",
	"sec_num": "4."
	},
	{
	"text": "The first two terms in the numerator are designed to account for the two types of intersection in 2 and 3, above. If some of the ILT filler falls outside the intersection, then S i -I > 0, and the first term will add to the penalty. If some of the GL filler lies outside of the intersection, the second term will add to the penalty. The determination of just how much is added in each case is dependent upon the weighting functions W i and W g . The penalty is also decreased by the size of the third term, which is the weighted intersection. (This counteracts the additive factor of I in the first two terms when I < 0.) Finally, the whole penalty is divided by the weighted size of the domain, which accounts for the varying probability of matching in larger domains.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "The fillers do not overlap.",
	"sec_num": "4."
	},
	{
	"text": "The matching of an unordered domain with enumerated filler types and of a continuous domain with a range filler type are also scored using equation 3.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Other Domain/Filler Types",
	"sec_num": "4.3.5"
	},
	{
	"text": "Once all of the slots have been compared using the appropriate equation, the overall score for the match between the ILT frame and the GL entry is produced, according to the following formula:",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "The Frame Level",
	"sec_num": "4.3.6"
	},
	{
	"text": "where the P i 's are the mismatch penalties between each slot of the ILT frame and the GL entry and the I i 's are the importance values for the corresponding slots of the GL entry meaning pattern. (The slots in an ILT frame do not contain importance values.)",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "The Frame Level",
	"sec_num": "4.3.6"
	},
	{
	"text": "The functions in equation 3 can be calibrated to yield results with an intuitive interpretation. For example, for most applications, the W g function should be quite low, since a subrange of a given range should reveal a valid instance of the concept. Conversely, the function W i should be quite high, signifying an entry which does not fit into the specified range. However, the values which are suitable for any particular application must be produced empirically.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Calibration",
	"sec_num": "4.3.7"
	},
	{
	"text": "In the DIOGENES generator, an instantiation of a head-selecting knowledge source is triggered simultaneously for every proposition and role instance in the input The results of their operation are posted to a blackboard (the Lexical Selection BlackBoard, LSBB), so that all knowledge source instances can draw on this knowledge in their own decision processes. The knowledge sources responsible for selecting modifiers are triggered when the heads of their phrases have already been selected. Figure 5 shows a simplified version of the top-level lexical selection algorithm. Extensions exist a) for the case when a change must be made in the lexical choice due to late evidence of collocational incompatibility, b) for selecting modifiers and c). for selecting definite descriptions in the anaphoric realization mode.",
	"cite_spans": [],
	"ref_spans": [
	{
	"start": 493,
	"end": 501,
	"text": "Figure 5",
	"ref_id": null
	}
	],
	"eq_spans": [],
	"section": "The Lexical Selection Algorithm",
	"sec_num": "5"
	},
	{
	"text": "The predicate anaphoric-realization-indicated returns t if an input frame is co-referential with a previously instantiated IBB unit and when relevant heuristic rules determine that it should be realized pronominally, elliptically or through a definite description.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "The Lexical Selection Algorithm",
	"sec_num": "5"
	},
	{
	"text": "The function do-anaphoric determines the appropriate realization as above and posts it on the LSBB. The function GL-search in Figure 6 searches the generation lexicon and produces a set of candidate lexical realizations by obtaining all entries that have the same is-token-of value that is found within the ILT frame.",
	"cite_spans": [],
	"ref_spans": [
	{
	"start": 126,
	"end": 134,
	"text": "Figure 6",
	"ref_id": "FIGREF3"
	}
	],
	"eq_spans": [],
	"section": "The Lexical Selection Algorithm",
	"sec_num": "5"
	},
	{
	"text": "It is clear that the acquisition of the generation lexicon is a major and extremely labor-intensive task. The acquisition of this dictionary, especially of the collocational information cannot at present be done automatically. But the efficiency of a team of lexicographers can be increased dramatically through the use of a specialized intelligent interactive aid, which will be both called through the interactively-augment-GL procedure during a generation session as well as being used for acquiring GL 'off-line.' We have developed one such Knowledge Acquisition Maintenance System (cf. Nirenburg et al., 1988) , called ONTOS, for the acquisition of domain knowledge and are at present extending it for acquiring generation lexicons as well. In the first months of running the generator in every new domain, we expect to use the on-line augmentation facility extensively, until all the relevant entries in GL are covered.",
	"cite_spans": [
	{
	"start": 591,
	"end": 614,
	"text": "Nirenburg et al., 1988)",
	"ref_id": "BIBREF21"
	}
	],
	"ref_spans": [],
	"eq_spans": [],
	"section": "The Lexical Selection Algorithm",
	"sec_num": "5"
	},
	{
	"text": "The function collocationally-constrain in Figure 7 first uses the function nunber-of-potential-collocations to find out whether a lexical choice has already been made in the current proposition that can render some of the members of the current candidate set inappropriate collocationally. Collocational constraints are checked between a) the proposition head and all the role heads, b) heads and modifiers and c) all conjoined concept instances in the input. If no potential collocations are found, the current process can either wait and check again later or exit collocationally-constrain without any filtering actually achieved. (If there is no such provision there is a danger of a deadlock when all the processes will be in the waiting mode.) This eventuality is taken care of by rigging the probabilities of the two outcomes in toss-a-coin 2 . Heuristically, the choice of the main verb (proposition head) is the most independent, followed by the choice of noun phrase heads. Collocationally-constrain uses the collocation information in the lexicon entries to match and filter the candidate sets.",
	"cite_spans": [],
	"ref_spans": [
	{
	"start": 42,
	"end": 50,
	"text": "Figure 7",
	"ref_id": "FIGREF4"
	}
	],
	"eq_spans": [],
	"section": "The Lexical Selection Algorithm",
	"sec_num": "5"
	},
	{
	"text": "If the residual set has cardinality one, the result is posted on the LSBB. If there exist more than one candidate, the function select-best (i.e., the Matcher) is called to perform context-independent lexical selection based on the quality of the match between the meaning pattern of the ILT frame on the one hand and the weighted meaning patterns of the GL entries in the candidate set on the other.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "The Lexical Selection Algorithm",
	"sec_num": "5"
	},
	{
	"text": "The predicate modifier-realization-indicated returns t if there exist properties in the ILT frame that were not accounted for in the head realization of the lexical item, i.e., realized as either a noun or a verb, roughly speaking (see Figure 5) .",
	"cite_spans": [],
	"ref_spans": [
	{
	"start": 236,
	"end": 245,
	"text": "Figure 5)",
	"ref_id": null
	}
	],
	"eq_spans": [],
	"section": "The Lexical Selection Algorithm",
	"sec_num": "5"
	},
	{
	"text": "The function select-modifier selects the appropriate modifier realization for the above phrase head, i.e., selecting an adjective for a noun and an adverb for a verb. This function uses essentially the same set of algorithms, i.e., GL-search, collocationally-constrain and select-best, for choosing phrase modifiers as was employed for selecting phrase heads. The primary difference between these sets of algorithms lies in the fact that the modifiers are constrained to collocate with only their respective phrase heads.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "The Lexical Selection Algorithm",
	"sec_num": "5"
	},
	{
	"text": "The blackboard architecture, the process handler and the inexact meaning matching module have been implemented; the collocation treatment module has also been implemented, but extensive testing has not been performed due to the lack of a large-scale lexicon. The anaphora treatment module has been implemented for pronominalization only. Our current tasks in the lexical selection module of DIOGENES thus include the acquisition of a sizable lexicon, complete with collocation information, extensions to the treatment of reference to incorporate definite descriptions, and the development of the module facilitating interaction between lexical and syntactic means of realization.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Status and Future Work",
	"sec_num": "6"
	},
	{
	"text": "Temperature is one of the physical object attributes within the ontological is-a hierarchy from which ILTs are produced(Nirenburg et al., 1987b).",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "",
	"sec_num": null
	},
	{
	"text": "This mechanism is similar to the one used in the Friendly-Neighbors algorithm of the PARPAR parallel parser(Lozinskii and Nirenburg, 1986).",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "",
	"sec_num": null
	}
	],
	"back_matter": [
	{
	"text": "slot-value := get-is-token-of (current-frame); candidate-set := first-pass-search (slot-value GL); how-many := cardinality (candidate-set) ;if how-many = 0 then candidate-set :=interactively-augment-GL (current-frame GL) else candidate-set := union (candidate-set how-many) ;post-intermediate-candidate-set (candidate-set) ;",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "annex",
	"sec_num": null
	}
	],
	"bib_entries": {
	"BIBREF0": {
	"ref_id": "b0",
	"title": "Planning English Sentences. ACL Series -Studies in Natural Language Processing",
	"authors": [
	{
	"first": "D",
	"middle": [
	"E"
	],
	"last": "Appelt",
	"suffix": ""
	}
	],
	"year": 1985,
	"venue": "",
	"volume": "",
	"issue": "",
	"pages": "",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Appelt, D.E. 1985. Planning English Sentences. ACL Series -Studies in Natural Language Processing. Cambridge: Cambridge University Press.",
	"links": null
	},
	"BIBREF1": {
	"ref_id": "b1",
	"title": "A Computational Model for Extemporaneous Elaborations",
	"authors": [
	{
	"first": "M",
	"middle": [
	"A"
	],
	"last": "Bienkowski",
	"suffix": ""
	}
	],
	"year": 1986,
	"venue": "CSL Report",
	"volume": "1",
	"issue": "",
	"pages": "",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Bienkowski, M.A. 1986. A Computational Model for Extemporaneous Elaborations. CSL Report 1, Cognitive Science Laboratory, Princeton University.",
	"links": null
	},
	"BIBREF2": {
	"ref_id": "b2",
	"title": "FRAMEKIT, A Reference Manual",
	"authors": [
	{
	"first": "J",
	"middle": [],
	"last": "Carbonell",
	"suffix": ""
	},
	{
	"first": "R",
	"middle": [],
	"last": "Joseph",
	"suffix": ""
	}
	],
	"year": 1985,
	"venue": "",
	"volume": "",
	"issue": "",
	"pages": "",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Carbonell, J. and R. Joseph. 1985. FRAMEKIT, A Reference Manual. Department of Computer Science. Carnegie-Mellon University.",
	"links": null
	},
	"BIBREF3": {
	"ref_id": "b3",
	"title": "The Lexicon in Text Generation. Paper presented at the LSA Linguistics Institute Workshop on Lexicon",
	"authors": [
	{
	"first": "S",
	"middle": [],
	"last": "Cumming",
	"suffix": ""
	}
	],
	"year": 1986,
	"venue": "",
	"volume": "",
	"issue": "",
	"pages": "",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Cumming, S. 1986. The Lexicon in Text Generation. Paper presented at the LSA Linguistics Institute Workshop on Lexicon, New York.",
	"links": null
	},
	"BIBREF4": {
	"ref_id": "b4",
	"title": "Conceptual and Linguistic Decisions in Generation",
	"authors": [
	{
	"first": "L",
	"middle": [],
	"last": "Danlos",
	"suffix": ""
	}
	],
	"year": 1984,
	"venue": "Proceedings of COLING-84",
	"volume": "",
	"issue": "",
	"pages": "",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Danlos, L. 1984. Conceptual and Linguistic Decisions in Generation. Proceedings of COLING-84.",
	"links": null
	},
	"BIBREF6": {
	"ref_id": "b6",
	"title": "Modes of Meaning",
	"authors": [
	{
	"first": "J",
	"middle": [
	"R"
	],
	"last": "Firth",
	"suffix": ""
	},
	{
	"first": "",
	"middle": [],
	"last": "London",
	"suffix": ""
	},
	{
	"first": "N",
	"middle": [],
	"last": "Goldman",
	"suffix": ""
	}
	],
	"year": 1951,
	"venue": "Conceptual Information Processing",
	"volume": "",
	"issue": "",
	"pages": "289--372",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Firth, J.R. 1951. Modes of Meaning. In: J.R. Firth, Papers in Linguistics, London. Goldman, N. 1975. Conceptual Generation. In: R. Schank (ed.), Conceptual Information Processing. Amsterdam: North Holland, pp. 289-372.",
	"links": null
	},
	"BIBREF7": {
	"ref_id": "b7",
	"title": "Semantic-Based Generation of Japanese-German Translation System",
	"authors": [
	{
	"first": "K",
	"middle": [],
	"last": "Hanakata",
	"suffix": ""
	},
	{
	"first": "A",
	"middle": [],
	"last": "Lesniewski",
	"suffix": ""
	},
	{
	"first": "S",
	"middle": [],
	"last": "Yokoyama",
	"suffix": ""
	}
	],
	"year": 1986,
	"venue": "Proceedings of COLING-86",
	"volume": "",
	"issue": "",
	"pages": "560--562",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Hanakata, K., A. Lesniewski, S. Yokoyama. 1986. Semantic-Based Generation of Japanese-German Translation System. Proceedings of COLING-86. Bonn, pp. 560-562.",
	"links": null
	},
	"BIBREF8": {
	"ref_id": "b8",
	"title": "Lexical Information for Parsing Systems: Points of Convergence and Divergence",
	"authors": [
	{
	"first": "R",
	"middle": [],
	"last": "Ingria",
	"suffix": ""
	}
	],
	"year": 1987,
	"venue": "",
	"volume": "",
	"issue": "",
	"pages": "",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Ingria, R. 1987. Lexical Information for Parsing Systems: Points of Convergence and Divergence. In: D. Walker, A. Zampolli and N. Calzolari (eds.), Automating the Lexicon: Research and Practice in a Multilingual Environment.",
	"links": null
	},
	"BIBREF9": {
	"ref_id": "b9",
	"title": "A Knowledge-Based Approach to Language Production",
	"authors": [
	{
	"first": "P",
	"middle": [],
	"last": "Jacobs",
	"suffix": ""
	}
	],
	"year": 1985,
	"venue": "",
	"volume": "",
	"issue": "",
	"pages": "",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Jacobs, P. 1985. A Knowledge-Based Approach to Language Production. Ph.D. Dissertation, University of California at Berkeley.",
	"links": null
	},
	"BIBREF10": {
	"ref_id": "b10",
	"title": "Metaphors We Live By",
	"authors": [
	{
	"first": "G",
	"middle": [],
	"last": "Lakoff",
	"suffix": ""
	},
	{
	"first": "M",
	"middle": [],
	"last": "Johnson",
	"suffix": ""
	}
	],
	"year": 1980,
	"venue": "",
	"volume": "",
	"issue": "",
	"pages": "",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Lakoff, G. and M. Johnson, 1980. Metaphors We Live By. Chicago: University of Chicago Press.",
	"links": null
	},
	"BIBREF11": {
	"ref_id": "b11",
	"title": "Language Generation from Conceptual Structure: Synthesis of German in a Japanese/German MT Project",
	"authors": [
	{
	"first": "J",
	"middle": [],
	"last": "Laubsch",
	"suffix": ""
	},
	{
	"first": "D",
	"middle": [],
	"last": "R\u00f6sner",
	"suffix": ""
	},
	{
	"first": "K",
	"middle": [],
	"last": "Hanakata",
	"suffix": ""
	},
	{
	"first": "A",
	"middle": [],
	"last": "Lesniewski",
	"suffix": ""
	}
	],
	"year": 1984,
	"venue": "Proceedings of COLING-84. Stanford",
	"volume": "",
	"issue": "",
	"pages": "491--494",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Laubsch, J., D. R\u00f6sner, K. Hanakata, and A. Lesniewski. 1984. Language Generation from Conceptual Structure: Synthesis of German in a Japanese/German MT Project. Proceedings of COLING-84. Stanford, pp. 491-494.",
	"links": null
	},
	"BIBREF12": {
	"ref_id": "b12",
	"title": "Parsing in Parallel",
	"authors": [
	{
	"first": "E",
	"middle": [
	"L"
	],
	"last": "Lozinskii",
	"suffix": ""
	},
	{
	"first": "S",
	"middle": [],
	"last": "Nirenburg",
	"suffix": ""
	}
	],
	"year": 1986,
	"venue": "Journal of Computer Languages",
	"volume": "11",
	"issue": "1",
	"pages": "39--51",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Lozinskii, E. L. and S. Nirenburg. 1986. Parsing in Parallel. Journal of Computer Languages, Volume 11, Number 1, pp. 39-51.",
	"links": null
	},
	"BIBREF13": {
	"ref_id": "b13",
	"title": "Nigel: a Systemic Grammar for Text Generation",
	"authors": [
	{
	"first": "W",
	"middle": [],
	"last": "Mann",
	"suffix": ""
	},
	{
	"first": "C",
	"middle": [
	"M I M"
	],
	"last": "Matthiessen",
	"suffix": ""
	}
	],
	"year": 1985,
	"venue": "Systemic Perspectives on Discourse",
	"volume": "",
	"issue": "",
	"pages": "",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Mann, W. and C.M.I.M. Matthiessen. 1985. Nigel: a Systemic Grammar for Text Generation. In: Freedle (ed.), Systemic Perspectives on Discourse. Norwood, NJ: Ablex.",
	"links": null
	},
	"BIBREF14": {
	"ref_id": "b14",
	"title": "Generation Systems Should Choose Their Words",
	"authors": [
	{
	"first": "M",
	"middle": [],
	"last": "Marcus",
	"suffix": ""
	}
	],
	"year": 1987,
	"venue": "Proceedings of TINLAP-3",
	"volume": "",
	"issue": "",
	"pages": "",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Marcus, M. 1987. Generation Systems Should Choose Their Words. Proceedings of TINLAP-3. Las Cruces, NM, February.",
	"links": null
	},
	"BIBREF15": {
	"ref_id": "b15",
	"title": "Natural Language Generation as a Computational Problem: an Introduction",
	"authors": [
	{
	"first": "D",
	"middle": [
	"D"
	],
	"last": "Mcdonald",
	"suffix": ""
	},
	{
	"first": "K",
	"middle": [],
	"last": "Mckeown",
	"suffix": ""
	}
	],
	"year": 1983,
	"venue": "Computational Problems in Discourse",
	"volume": "",
	"issue": "",
	"pages": "",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "McDonald, D. D. 1983. Natural Language Generation as a Computational Problem: an Introduction. In: Brady and Berwick (eds.), Computational Problems in Discourse. Cambridge: MIT Press. McKeown, K. 1985. Text Generation. Cambridge: Cambridge University Press.",
	"links": null
	},
	"BIBREF16": {
	"ref_id": "b16",
	"title": "Towards a Theory of Linguistic Models of the Meaning-Text Type",
	"authors": [
	{
	"first": "L",
	"middle": [
	"A"
	],
	"last": "Mel'\u010duk",
	"suffix": ""
	}
	],
	"year": 1974,
	"venue": "",
	"volume": "",
	"issue": "",
	"pages": "",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Mel'\u010duk, LA. 1974. Towards a Theory of Linguistic Models of the Meaning-Text Type. Moscow: Nauka.",
	"links": null
	},
	"BIBREF17": {
	"ref_id": "b17",
	"title": "Meaning -Text Models: A Recent Trend in Soviet Linguistics",
	"authors": [
	{
	"first": "I",
	"middle": [
	"A"
	],
	"last": "Mel'\u010duk",
	"suffix": ""
	}
	],
	"year": 1981,
	"venue": "Annual Review of Anthropology",
	"volume": "10",
	"issue": "",
	"pages": "27--62",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Mel'\u010duk, I.A. 1981. Meaning -Text Models: A Recent Trend in Soviet Linguistics. Annual Review of Anthropology. Vol. 10, pp. 27-62.",
	"links": null
	},
	"BIBREF18": {
	"ref_id": "b18",
	"title": "On Knowledge-Based Machine Translation",
	"authors": [
	{
	"first": "S",
	"middle": [],
	"last": "Nirenburg",
	"suffix": ""
	},
	{
	"first": "V",
	"middle": [],
	"last": "Raskin",
	"suffix": ""
	},
	{
	"first": "A",
	"middle": [],
	"last": "Tucker",
	"suffix": ""
	}
	],
	"year": 1986,
	"venue": "Proceedings of COLING-86",
	"volume": "",
	"issue": "",
	"pages": "627--632",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Nirenburg, S., V.Raskin and A. Tucker. 1986. On Knowledge-Based Machine Translation. Proceedings of COLING-86. Bonn, pp. 627-632.",
	"links": null
	},
	"BIBREF19": {
	"ref_id": "b19",
	"title": "Integrating Discourse Pragmatics and Propositional Knowledge for Multi-Lingual Natural Language Processing",
	"authors": [
	{
	"first": "S",
	"middle": [],
	"last": "Nirenburg",
	"suffix": ""
	},
	{
	"first": ";",
	"middle": [
	"S"
	],
	"last": "",
	"suffix": ""
	},
	{
	"first": "J",
	"middle": [],
	"last": "Carbonell",
	"suffix": ""
	}
	],
	"year": 1987,
	"venue": "Computers and Translation",
	"volume": "",
	"issue": "2",
	"pages": "105--116",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Nirenburg, S. 1987. A Distributed System for Language Generation. Technical Report CMU-CMT-87- 102. Carnegie-Mellon University. May. Nirenburg, S. and J. Carbonell. 1987. Integrating Discourse Pragmatics and Propositional Knowledge for Multi-Lingual Natural Language Processing. In: Computers and Translation, Number 2, pp. 105-116. Nirenburg, S., E. Nyberg, and E. Kenschaft. 1987a. Inexact Frame Matching for Lexical Selection in Natural Language Generation. CMU-CMT Unpublished Memo.",
	"links": null
	},
	"BIBREF20": {
	"ref_id": "b20",
	"title": "The Structure of Interlingua in TRANSLATOR",
	"authors": [
	{
	"first": "S",
	"middle": [],
	"last": "Nirenburg",
	"suffix": ""
	},
	{
	"first": "V",
	"middle": [],
	"last": "Raskin",
	"suffix": ""
	},
	{
	"first": "A",
	"middle": [],
	"last": "Tucker",
	"suffix": ""
	}
	],
	"year": 1987,
	"venue": "",
	"volume": "",
	"issue": "",
	"pages": "",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Nirenburg, S., V. Raskin and A. Tucker. 1987b. The Structure of Interlingua in TRANSLATOR. In: S.",
	"links": null
	},
	"BIBREF21": {
	"ref_id": "b21",
	"title": "Acquisition of Very Large Knowledge Bases: Methodology, Tools and Applications. Submitted to the Second International Workshop on Knowledge Acquisition",
	"authors": [
	{
	"first": "S",
	"middle": [],
	"last": "Nirenburg ; Nirenburg",
	"suffix": ""
	},
	{
	"first": "I",
	"middle": [],
	"last": "Monarch",
	"suffix": ""
	},
	{
	"first": "T",
	"middle": [],
	"last": "Kaufmann",
	"suffix": ""
	},
	{
	"first": "I",
	"middle": [],
	"last": "Nirenburg",
	"suffix": ""
	},
	{
	"first": "J",
	"middle": [],
	"last": "Carbonell",
	"suffix": ""
	}
	],
	"year": 1988,
	"venue": "Machine Translation: Theoretical and Methodological Issues, ACL Series -Studies in Natural Language Processing",
	"volume": "",
	"issue": "",
	"pages": "90--113",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Nirenburg (ed.), Machine Translation: Theoretical and Methodological Issues, ACL Series -Studies in Natural Language Processing. Cambridge: Cambridge University Press, pp. 90-113. Nirenburg, S., I. Monarch, T. Kaufmann, I. Nirenburg and J. Carbonell. 1988. Acquisition of Very Large Knowledge Bases: Methodology, Tools and Applications. Submitted to the Second International Workshop on Knowledge Acquisition. Banff, Canada. Nyberg, E. 1988. The FrameKit User's Guide. CMU-CMT Memo.",
	"links": null
	},
	"BIBREF22": {
	"ref_id": "b22",
	"title": "When Mariko Talks to Siegfried",
	"authors": [
	{
	"first": "D",
	"middle": [],
	"last": "R\u00f6sner",
	"suffix": ""
	}
	],
	"year": 1986,
	"venue": "Proceedings of COLING-86",
	"volume": "",
	"issue": "",
	"pages": "652--654",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "R\u00f6sner, D. 1986. When Mariko Talks to Siegfried. Proceedings of COLING-86. Bonn, pp. 652-654. Ward, N. 1988. Issues in Word Choice. To appear in Proceedings of COLING-88, Budapest.",
	"links": null
	}
	},
	"ref_entries": {
	"FIGREF0": {
	"num": null,
	"uris": null,
	"text": "<value>) (importance <importance-value>))] ) <CL-concept> ::= {any concept in the Concept Lexicon} <relation> ::= {any relation from the Concept Lexicon} <value> ::= {any concept or attribute (scale) value in the Concept Lexicon} <importance-value> ::= 1 \| 2 \| ... \| 10 <TL-pattern> ::= (<TL-lexeme> <lex-info> <collocation> ) <TL-lexeme> ::= (<language> TL-lexical-unit \| (synonym TL-lexical-unit)) <language> ::= english \| spanish \| russian \| japanese \| ... <lex-info> ::= ((<syntactic-info>) (morph <inflection-type>)) <syntactic-info> ::= {the contents of a syntactic dictionary (cf. e.g. Ingria, 1987)} <inflection-type> :: = {an indication of irregularities in forming word forms, e.g., goose -pl. geese} <collocation> ::= ( {<dimension> <dimension-value>} * ) <dimension> ::= {the name of a lexical collocation relation based on the Concept Lexicon slot names for the concept in question} <dimension-value> ::= {a TL lexical unit (word or expression) that can ordinarily collocate with the TL lexical unit in <TL-lexeme> above and connected to the TL unit on a specified dimension; can be recursive} Figure 1: The Structure of the Generation Lexicon",
	"type_str": "figure"
	},
	"FIGREF1": {
	"num": null,
	"uris": null,
	"text": "The GL Meaning Pattern for girl",
	"type_str": "figure"
	},
	"FIGREF2": {
	"num": null,
	"uris": null,
	"text": "",
	"type_str": "figure"
	},
	"FIGREF3": {
	"num": null,
	"uris": null,
	"text": "The GL-search Algorithm function collocationally-constrain (candidate-set) ; if number-of-potential-collocations (current-frame IBB) <> 0 then collocationally-constrain := intersection (candidate-set neighbor.collocation.!frame) else begin If toss-a-coin (current-frame) = heads then try-later else collocationally-constrain := candidate-set; end; post-intermediate-candidate-set (candidate-set) ;",
	"type_str": "figure"
	},
	"FIGREF4": {
	"num": null,
	"uris": null,
	"text": "The collocationally-constrain Algorithm",
	"type_str": "figure"
	},
	"TABREF0": {
	"text": "\u2022 C: the size of the domain \u2022 W i : a weighting function dependent upon the ILT filler size \u2022 W g : a weighting function dependent upon the GL filler size \u2022 W int : a weighting function dependent upon the size of the intersection",
	"num": null,
	"content": "<table><tr><td>\u2022 F(C): a damping function dependent upon the domain size</td></tr></table>",
	"type_str": "table",
	"html": null
	}
	}
	}
	}