Datasets:

WINGNUS
/

ACL-OCL

	{
	"paper_id": "W94-0309",
	"header": {
	"generated_with": "S2ORC 1.0.0",
	"date_generated": "2023-01-19T04:47:09.575217Z"
	},
	"title": "ON MOVING ON ON ONTOLOGIES: MASS, COUNT AND LONG THIN THINGS",
	"authors": [
	{
	"first": "Robin",
	"middle": [
	"P"
	],
	"last": "Fawcett",
	"suffix": "",
	"affiliation": {
	"laboratory": "Computational Linguistics Unit",
	"institution": "University of Wales",
	"location": {
	"postCode": "CF1 3EU",
	"settlement": "Cardiff",
	"country": "UK"
	}
	},
	"email": "fawcett@cardiff.ac.uk"
	}
	],
	"year": "",
	"venue": null,
	"identifiers": {},
	"abstract": "This paper discusses the principles that should govern the construction of two components of a system for natural language generation (NLG): (1) the ontology-or, rather, as the paper argues, the 'ontological' aspects of a belief system-and (2) the semantic representation of noun senses. It is an interesting fact that many ontologies bear a striking resemblance to a system network, as used in systemic functional grammar (SFG). Furthermore, two major current research efforts in the field of ontologybuilding are designed to run with a SFG generator: Pangloss, where the generator is Penman, and COMMUNAL, where the generator is GENESYS. It is therefore important to establish a principled approach to the 'division of labour' between the ontology and the equivalent aspects of the model of language-here a system network for the 'meaning potential' of English nouns. (However, the general principles should be relevant to ANY model of language.) The paper summafises (a)the purposes and (b)the structure of (1) a system network for noun senses and (2) the equivalent ontology (based on what we in the COMMUNAL Project judge is required in the next generation of belief systems for NLG). Examples are given of current work on the relevant system network and, more briefly, of the equivalent ontological aspects of the belief system. In particular, reasons are given why it would be inappropriate to give a primary place to the 'mass' vs. 'count' distinction in an 'interlingua' ontology-and even, surprising though it may seem, in a language-specific semantics for English. Finally, it turns out that, in the new perspective presented here, there is no 'component' of the belief system that is 'the ontology', and the reasons for this apparently anomalous position are given.",
	"pdf_parse": {
	"paper_id": "W94-0309",
	"_pdf_hash": "",
	"abstract": [
	{
	"text": "This paper discusses the principles that should govern the construction of two components of a system for natural language generation (NLG): (1) the ontology-or, rather, as the paper argues, the 'ontological' aspects of a belief system-and (2) the semantic representation of noun senses. It is an interesting fact that many ontologies bear a striking resemblance to a system network, as used in systemic functional grammar (SFG). Furthermore, two major current research efforts in the field of ontologybuilding are designed to run with a SFG generator: Pangloss, where the generator is Penman, and COMMUNAL, where the generator is GENESYS. It is therefore important to establish a principled approach to the 'division of labour' between the ontology and the equivalent aspects of the model of language-here a system network for the 'meaning potential' of English nouns. (However, the general principles should be relevant to ANY model of language.) The paper summafises (a)the purposes and (b)the structure of (1) a system network for noun senses and (2) the equivalent ontology (based on what we in the COMMUNAL Project judge is required in the next generation of belief systems for NLG). Examples are given of current work on the relevant system network and, more briefly, of the equivalent ontological aspects of the belief system. In particular, reasons are given why it would be inappropriate to give a primary place to the 'mass' vs. 'count' distinction in an 'interlingua' ontology-and even, surprising though it may seem, in a language-specific semantics for English. Finally, it turns out that, in the new perspective presented here, there is no 'component' of the belief system that is 'the ontology', and the reasons for this apparently anomalous position are given.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Abstract",
	"sec_num": null
	}
	],
	"body_text": [
	{
	"text": "'Or\" I One of the =lvens of Computational Linguistics (CL)which is taken here in a sense that includes Machine Translation (MT) -is that any such system needs an ontology, l But what, precisely, is an ontology? It seems to be one of those concepts which everyone who works with it instinctively feels tfiey understand, so that the basic assumptions are seldom made explicit. In practical terms, there is a fairly general assumption that an ontology is closely related to, and perhaps isomorphic with, the 'meanings' of the nouns of a language -or of a set of languages, the maximal set being all human languages. But in buildin,q a theoretically satisfactory overall model one discovers tI~at there are serious problems with this position, as will be shown in this paper.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Some current issues in modelling 'ontologies'",
	"sec_num": "1"
	},
	{
	"text": "There is a long history of work on ontolo,~ies for CL, including/he important work at Carnegie-Meffbn University over many years, that of Dahlgren and her collea,~ues (e.g. Dahlgren 1988 ) and the current Pangloss ProjecT, as described in Hovy and Nirenburg 1992 , Hovy and Knight 1993 . We in the COMMUNAL Project have been considering alternative approaches to this aspect of what we term the belief system, and I would like to present here, for discussion by the wider NLG community, the principles that we have established, often after years of experimentation, as they relate to these matters. We are currently implementing a system based on these principles.",
	"cite_spans": [
	{
	"start": 138,
	"end": 186,
	"text": "Dahlgren and her collea,~ues (e.g. Dahlgren 1988",
	"ref_id": null
	},
	{
	"start": 239,
	"end": 262,
	"text": "Hovy and Nirenburg 1992",
	"ref_id": "BIBREF14"
	},
	{
	"start": 263,
	"end": 285,
	"text": ", Hovy and Knight 1993",
	"ref_id": "BIBREF13"
	}
	],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Some current issues in modelling 'ontologies'",
	"sec_num": "1"
	},
	{
	"text": "In many respects, of course, our assumptions are similar to those of others working in this area. But our view is that the next generation of systems in Artificial Intelligence (AI) -and possibly also in MT -will require, as central components, belief systems that represent knowledge (or, more accurately, beliefs) of more types and in a more complex manner than in some current systems. As we shall see by the end of this paper, the phenomena that are often handled in terms of an ontolo~ov look somewhat different in this new perspective. In relati~ia to some of the issues to be discussed here, then, we are constructing a different overall model from that which appears to underlie much other current work. The purpose of this paper is to set out these ideas, to compare them with those of other researchers on whose work we are seeking to build, and to give some explanation of why we are following the direction that we are. We hope that this will open up further discussion about the next generation of belief systems.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Some current issues in modelling 'ontologies'",
	"sec_num": "1"
	},
	{
	"text": ". . The first step is to be clear about what the issues are. They are (1) issues of levels (which we shall here assume to mean levels of language), (2) issues of components of the overall system, and (3) issues of the structure and contents of part of the largest component of the overall system, namely the belief system. We shall focusparticularly on the types of relations that need to be recogmsed as holding between the 'concepts' in an ontology or rather, between the gene,ric, objects (in contrast with specific objects), such as dog, realized as dogs, as in I like dogs.2 Hovy and Nirenburg (1992) , in clearing the ground for their discussion of the principles that should guide the construction of an ontology, suggest that most ontologies and domain models to date have been assembled based primarily on introspection, and often reflect the idiosyncrasies of the builders more than the requirements of the application (such as MT). Lacking well-founded guided principles, the ontology builder is working in the dark.",
	"cite_spans": [
	{
	"start": 580,
	"end": 605,
	"text": "Hovy and Nirenburg (1992)",
	"ref_id": "BIBREF14"
	}
	],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Some current issues in modelling 'ontologies'",
	"sec_num": "1"
	},
	{
	"text": "This judgement seems a little hyperbolic, in view of the fact that a number of recent ontology-builders have explained their principles as fully as Hovy and Nirenburg do. Thus, while there are aspects of Dahlgren s frameworI( (1988:46f.) that are open to criticism, she in fact gives as detailed an account of her principles as do Hovy and Nirenburg. Bateman (1990, Bate, man et al 19,90 ) similarly explains the ideas underlyin,q the uover model in Penman. In fact, as Hovy and KnigSht (19~) state, the 'ontology base' in Pangloss is in part based on Penman (being a merger of this, the semantic categories from the Longman Dictionary of Contemporary English (LDOCE) , which is intended as a taxonomy for the nouns of English), and ULTRA (Nirenburg and Defrise 1992) -which itself draws on the LDOCE categories). Nonetheless the main thrust of 7th International Generation Workshop \u2022 Kennebunkport, Maine \u00b0 June [21] [22] [23] [24] 1994 Hovy and Nirenberg's comment is surely right, in that the discussion of the principles governing the building onto!ogies is far from over. Indeed, it may be that the reqmrements of MT and AI (where NL is used to communicate with a problem solving system) are different (at least at this stage of the development of MT).",
	"cite_spans": [
	{
	"start": 204,
	"end": 237,
	"text": "Dahlgren s frameworI( (1988:46f.)",
	"ref_id": null
	},
	{
	"start": 331,
	"end": 387,
	"text": "Hovy and Nirenburg. Bateman (1990, Bate, man et al 19,90",
	"ref_id": null
	},
	{
	"start": 470,
	"end": 492,
	"text": "Hovy and KnigSht (19~)",
	"ref_id": null
	},
	{
	"start": 660,
	"end": 667,
	"text": "(LDOCE)",
	"ref_id": null
	},
	{
	"start": 739,
	"end": 767,
	"text": "(Nirenburg and Defrise 1992)",
	"ref_id": "BIBREF22"
	},
	{
	"start": 913,
	"end": 917,
	"text": "[21]",
	"ref_id": null
	},
	{
	"start": 918,
	"end": 922,
	"text": "[22]",
	"ref_id": null
	},
	{
	"start": 923,
	"end": 927,
	"text": "[23]",
	"ref_id": null
	},
	{
	"start": 928,
	"end": 932,
	"text": "[24]",
	"ref_id": null
	},
	{
	"start": 933,
	"end": 937,
	"text": "1994",
	"ref_id": null
	}
	],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Why the discussion remains open",
	"sec_num": "2"
	},
	{
	"text": "Most -and perhaps all -ontology-builders accept that they are in fact working on the basis of natural languageand, typically, that thelanguage is English. In Bateman's words (1990:56) 'the Upper Model represents the speaker's ,e, xpenence in terms of ,generalized lingu!sfically-motivated ontological\" categories. He states that it is essential that constraints should be found for what an upper model should contain and how it should be organised [his italics], and he then goes on to suggest that it is the aspects of meaning found in the experiential meta-function in a systemic functional grammar that should guide the construction of the ontology. This application of SFG concepts provides a helpful framework in which to approach the problem. (We might note that distinctions that depend on register, e.g. on tenor (formality), such as that between cigarette and fag are irrelevant to the ontology.) The reason for depending on NL, then, seems to be that, without it, we would have no guidelines as to how to structure the ontology.",
	"cite_spans": [
	{
	"start": 158,
	"end": 183,
	"text": "Bateman's words (1990:56)",
	"ref_id": null
	}
	],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Why the discussion remains open",
	"sec_num": "2"
	},
	{
	"text": "Let us assume for the moment that this position is justified (though I shall suggest some problems with it in sections 5 and 6). This brings us to an important question for those who use the Upper Model or any derived ontology such as Pangloss. This is: if there, is to be this strong connection to the linguistic system (Bateman 1990:57) , what part of it should that connection be to?",
	"cite_spans": [
	{
	"start": 321,
	"end": 338,
	"text": "(Bateman 1990:57)",
	"ref_id": null
	}
	],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Why the discussion remains open",
	"sec_num": "2"
	},
	{
	"text": "And here we come to an apparentproblem for this line of argument. This is that the Nigel-model of language around which Penman is constructed lacks a specific network for noun senses. So what are the principles on which these aspects of the Upper Model are constructed? We shall return to this matter after the next section.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Why the discussion remains open",
	"sec_num": "2"
	},
	{
	"text": "Two extraneous factors that may lead to differing research assumptions First, however, let us be explicit about two of the factors that differentiate various research projects in this area, and so, perhaps, the assumptions underlying various conceptual frameworks that have been adopted. The first is the disciplinary coverage of ,researchers on a projectincluding the 'home discipline, as it were, of the leading researcher. Those such as myself whose starting point was linguistics are sometimes shocked at the cavalier way (note the loaded language to express the viewpoint!) in which nonlinguists simply adopt the 'senses' of the nouns of English as the starting point for an ontology. AI-minded computer scientists may be equally shocked at the pussyfooting way in which many linguists refuse to recognise the need to move outside the semiotic system of language, even though this is patently necessary in order to budd adequate models of how language is used. Often neither really addresses the issue of where the semiotic system of a natural language ends (i.e. the semantics of the language) and where the ',concepts,' (or whatever is assumed as the category for thinking ) begins. In this paper I shall set out a clear and, I believe, defensible position on these issues.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "3",
	"sec_num": null
	},
	{
	"text": "A second factor which undoubtedly affects the conceptual framework used in any given research project is the time scale set by the sponsors of one's research. Practically all sponsors of CL research expect products that are at least potentially 'applicable' (though often in a sense that is not well defined). For some researchers the time scale may be such that they must work with existing data bases, such as the machine readable version of the Longman Dictionary of Contemporary English (LDOCE) (Longman Group 1978) or 'Wordnet' (Miller 1990) . This seems to be the case with the Pangloss project (Hovy and Nirenburg 1992 , Hovy and Knight 1993 , Knight 1993 , whose explicit goal is to combine the best features of these sources. This is an ambitious goal and I wish the researchers well. However the wen-known fact must be pointed out that, in the last decade and a half, many other researchers have put in a lot of work in trying to make LDOCE usable in a number of ways -and yet so far as I am aware no one yet has found a way to use these data as part of a belief system without an enormous amount of handediting. There are important questions that need to be asked about the relationship of these data to ontology building. The answers should relate to an integrated framework that provictes appropriately for at least the two linguistic levels of meaning and form, and, outside language, the categories of a belief system ('concepts').",
	"cite_spans": [
	{
	"start": 499,
	"end": 519,
	"text": "(Longman Group 1978)",
	"ref_id": null
	},
	{
	"start": 533,
	"end": 546,
	"text": "(Miller 1990)",
	"ref_id": "BIBREF21"
	},
	{
	"start": 601,
	"end": 625,
	"text": "(Hovy and Nirenburg 1992",
	"ref_id": "BIBREF14"
	},
	{
	"start": 626,
	"end": 648,
	"text": ", Hovy and Knight 1993",
	"ref_id": "BIBREF13"
	},
	{
	"start": 649,
	"end": 662,
	"text": ", Knight 1993",
	"ref_id": "BIBREF15"
	}
	],
	"ref_spans": [],
	"eq_spans": [],
	"section": "3",
	"sec_num": null
	},
	{
	"text": "Researchers who are workingto a less constrained timetable are perhaps more able to ask such questions. There are arguments for and against each approach, and it is not the purpose of this paper to criticize the work of those who seek directly to exploit existing data bases. Indeed, it may be that such work will in time produce solutions to the problems to be discussed here, by developing EVOLUTION-ARILY into more advanced models. Alternatively, it may be that a significantly different framework is required in order to achieve optimal representations of belief systems; both lines of inquiry should be pursued.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "3",
	"sec_num": null
	},
	{
	"text": "In the COMMUNAL Project our task is to think speculatively about the next generation of belief systems, and about the components, relations and procedures that will be required in it. It is in the nature of research that we shall almost certainly have overlooked some aspects that will strike future researchers as important, but the enterprise is nonetheless worth attempting. Here, however, we shall not tU/to provide a complete overview, even very briefly, of all ot the components that we believe to be necessary in a belief system (for which see Fawcett 1993) , but just those aspects that pertain to the concept of the 'ontology'.",
	"cite_spans": [
	{
	"start": 551,
	"end": 564,
	"text": "Fawcett 1993)",
	"ref_id": "BIBREF8"
	}
	],
	"ref_spans": [],
	"eq_spans": [],
	"section": "3",
	"sec_num": null
	},
	{
	"text": "4 The intertwined concepts of 'ontology' and 'system network': a brief history",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "3",
	"sec_num": null
	},
	{
	"text": "Since ontologies display many of the characteristics of a s,~stem network, let us begin with Halliday's seminal paper Categories of the theory of grammar (1961) . In it he proposed the concept of lexis 'as most delicate grammar'. He envisaged a model of language in which the earlier choices in a system network would be realised grammatically, i.e. in the structures of clause and group syntax and in grammatical items -and where these earlier choices would lead on to more 'delicate' choices that would be realised as iexical items.",
	"cite_spans": [
	{
	"start": 154,
	"end": 160,
	"text": "(1961)",
	"ref_id": null
	}
	],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Halliday's proposal as a starting point",
	"sec_num": "4.1"
	},
	{
	"text": "In the following decades Halliday and others did a great deal of work to develop the grammatical aspects of the model -including the important step of integrating intonat!on with grammar; But what of integrating lexis~. While grammatical items such as modal verbs and various types o~ determiner were modelled in system networks, the concept of system networks for lexical items lexis remained largely unexplored until the mid-70s to the mid-80s. In that penoa there were several small studies by systemic linguists (though not by Halliday himself) which implemented the concept of lexis in networks (Berry 1977 , Fawcett 1980 , Hasan 1987 ), but they were simply illustrative and there was no attempt to explore the implications of a comprehensive treatment of the original concept. (These 'implementations' were linguistic descriptions, not computer implementations.)",
	"cite_spans": [
	{
	"start": 600,
	"end": 611,
	"text": "(Berry 1977",
	"ref_id": "BIBREF2"
	},
	{
	"start": 612,
	"end": 626,
	"text": ", Fawcett 1980",
	"ref_id": "BIBREF6"
	},
	{
	"start": 627,
	"end": 639,
	"text": ", Hasan 1987",
	"ref_id": "BIBREF12"
	}
	],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Halliday's proposal as a starting point",
	"sec_num": "4.1"
	},
	{
	"text": "Meanwhde, Halliday had added the important notion of explicitly proposed as networks of features intended to capture the meaning potential of a language. It has always seemed clear to me that, since the networks specify meaning potential, the features in them can be appropriately regardedas semantic features (Fawcett 1973 (Fawcett /81, 1980 . Halliday himself sometimes seems to agree with this position and sometimes to claim that the networks are at the level of form. (See the discussions of his variable position in Butler 1985 and , for example, Fawcett, Tucker and Lin 1993.) For the fullest and best account of the developments in SFG in modelling lexis, we must await Tucke?s PhD thesis, currently in preparation.",
	"cite_spans": [
	{
	"start": 310,
	"end": 323,
	"text": "(Fawcett 1973",
	"ref_id": "BIBREF5"
	},
	{
	"start": 324,
	"end": 342,
	"text": "(Fawcett /81, 1980",
	"ref_id": null
	},
	{
	"start": 522,
	"end": 537,
	"text": "Butler 1985 and",
	"ref_id": "BIBREF3"
	},
	{
	"start": 538,
	"end": 572,
	"text": ", for example, Fawcett, Tucker and",
	"ref_id": null
	},
	{
	"start": 573,
	"end": 583,
	"text": "Lin 1993.)",
	"ref_id": "BIBREF10"
	}
	],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Halliday's proposal as a starting point",
	"sec_num": "4.1"
	},
	{
	"text": "To the best of my belief, the first example of the intertwining of the two concepts ot an ontology and a system network occurred in the mid 1960s. It was at about this time that the system networks of SFG were beginning to be semanticized as the meaning potential of a language. Linguists working in the Chomskyan tradition had recently introduced the concept of selectional restrictions -which effectively presuppose semantic features. At this juncture a British hnguist, Geoffrey Leech, who was at University College London at the same time as Halliday was developing SFG there in the 1970s, made the interesting experiment of combining the two concepts of system networks and semantic features in his Towards a Semantic Description of English, published in 1969 (and later incorporated in his standard text book Semantics (Leech 1974/81 ). This, then, was an early attempt to provide an ontology-based logic for reasoning, and it was done at a level that Leech assumed to be the level of semantics, i.e., presumably, a level within language. (In the view taken in this paper and to be expanded later, reasoning is i'n fact better modelled as taking place at a level outside and above' language -while being heavily influenced by languag,e.) Leecn's model included a taxonomy of 'types of object, and it had most of the characteristics of current ontologies (which we shall summarise in Section 6). This work was, in effect, Leech's attempt to combine the relevant parts of a system network with the demands of a reasoning system. It was related to a fairly standard logic, so that the taxonomy could be used for simple reasoning tasks, in some of the ways that current ontologies,are expected to.3 The crucial point, however, is that Leech s network was not in fact a system network, as the term is used in SFG, but an ontology.",
	"cite_spans": [
	{
	"start": 825,
	"end": 839,
	"text": "(Leech 1974/81",
	"ref_id": null
	}
	],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Leech's logical semantics",
	"sec_num": "4.2"
	},
	{
	"text": "One striking v, isual icon of, this difference is the contrast between Halliday s and Leech s notations. They both use the usual systemi,c ,notation of a right-opening square bracket to ,indicate or and a curly right-opening bracket to mean and. But in Halliday's diagrams there is always an arrow pointing right, i.e. from the term (or terms) that is/are the entry condi,tion to the system towards the system itself. And in Leech s diagrams the arrows point from the system to the entry condition. In other words, the two apparently similar structures are to be used in different ways. As we shall see in Section 6, an ontology is typically (but not necessarily exclusively) traversed from right to left, which explains the direction of Leech's arrows. But a system network is intended to be traversed from left to ,right, 1\"., e. from the less 'delicate' choices to the more delicate ; see further below. (Some later systemic linguists, including myself, have followed Winograd (1972) in dispensing with the arrow. 4)",
	"cite_spans": [
	{
	"start": 971,
	"end": 986,
	"text": "Winograd (1972)",
	"ref_id": "BIBREF27"
	}
	],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Leech's logical semantics",
	"sec_num": "4.2"
	},
	{
	"text": "Another ontology with interesting similarities to a system network is found in the important work of Dahl~ren (1988) .",
	"cite_spans": [
	{
	"start": 110,
	"end": 116,
	"text": "(1988)",
	"ref_id": null
	}
	],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Dahlgren's ontology",
	"sec_num": "4.3"
	},
	{
	"text": "Her description of her 'category cuts' caff be expressed directly as a system network, with 'entity as the entry condition that leads immediately to two simultaneous Stems; 'individual' vs. 'collective' and 'abstract' vs. 'real'. rther del:~ndent systems, some of which are entered simultane, ously (i.e. as 'cross-classifications'), introduce a total of 37 category cuts' which, when all possible combinations are counted, generate 4272 potential ~combinations'. (This assumes that 'collective' leads on to 'mass' vs. 'set' vs. 'structure', as implied at one point; at another it does not.)",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Dahlgren's ontology",
	"sec_num": "4.3"
	},
	{
	"text": "Interestingly, Dahlgren advises at one point that 'care must be taken when adding cross-classificatmns at a hi,gher node ..... to avoid proliferating empty terminal nodes .....",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Dahlgren's ontology",
	"sec_num": "4.3"
	},
	{
	"text": "[because this] is a sur, e sign that the proposed crossclassification is spurious. If this were indeed to be accepted as a major critenon, one would unfortunately have to give her ontology rather low marks. This is because, of her 4272 'terminal nodes', most are unused (all but 57 on one count).5 In fact, as we shall see in Section 5, Dahlgren's advice is fully appropriate for a systemic linguist who is using the network as a generator -but it is certainly much less relevant to an ontology. Why should this be so.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Dahlgren's ontology",
	"sec_num": "4.3"
	},
	{
	"text": "-9",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Dahlgren's ontology",
	"sec_num": "4.3"
	},
	{
	"text": "The answer is that Dahlgren's taxonomy is not intended as a system network, and so it should not be thought of as operating by being traversed from left to right (e.g. as if to generate noun senses). An ontology is in fact typically used deductively, i.e. working from right to left TO REASON ABOUT OBJECTS (as Leech's arrows remind us). So the supposed 'overgeneration' of Dahlgren's ontology has no practical consequence -unless one wishes to use the ontology inductively. (In other words, many systems regularI~, use reasoning of the nature of 'If X is a rose then X is a flower, while there seems to be less demand for inductive reasoning. We shall return to the topic of reasoning in Section 6.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Dahlgren's ontology",
	"sec_num": "4.3"
	},
	{
	"text": "I cite this case in order to show the need for clear principles in the construction of both (1) system networks at the level of semantics, i.e. within language, and (2) ontological relations in a belief system.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Dahlgren's ontology",
	"sec_num": "4.3"
	},
	{
	"text": "We turn now to natural language generation systems that are based on SFG. There have been many of these, and many others that have incorporated significant aspects of the SFG approach to language (see Matthiessen and Bateman 1991 and Fawcett, Tucker and Lin 1993 for overviews of the use of SFGs in NLG). Here we shall consider just the two major SFG natural language generators: Penman, which is to be used as the generator for the current Pangloss project, and GENESYS, which is the generator in the COMMUNAL Project. Given the significant place of SFG generators in NLG, it is important to be clear about the theoretical framework for what such enterprises are attempting. In particular, it is important to understand the relation between the system network-and the ontology.",
	"cite_spans": [
	{
	"start": 201,
	"end": 216,
	"text": "Matthiessen and",
	"ref_id": "BIBREF19"
	},
	{
	"start": 217,
	"end": 233,
	"text": "Bateman 1991 and",
	"ref_id": "BIBREF19"
	},
	{
	"start": 234,
	"end": 253,
	"text": "Fawcett, Tucker and",
	"ref_id": "BIBREF10"
	},
	{
	"start": 254,
	"end": 262,
	"text": "Lin 1993",
	"ref_id": "BIBREF10"
	}
	],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Dahlgren's ontology",
	"sec_num": "4.3"
	},
	{
	"text": "The Penman Project was the first lar,~e SFG-based generator, and the mare work that establishes the structure and nature of what Halliday has called the lexieogrammar was done in the very early eighties (based fairly closely on Halliday's work of the seventms). In that period there was much more emphasis in linguistics as a whole on syntax and much less on lexis than there is now, and Halhday's programmatic use of the term 'lexicogrammar' was a farsighted pointer to where work would be needed in the future. But Halliday himself has always approached language from the grammatical rather than the lexmal end and, under his guidance, Mann and Matthiessen naturally worked first on the grammatical structures and items. Unfortunately, when the time came to extend the model to lexis, the sponsors of the project (working within the traditional framework of the 'grammar-vocabulary' distinction) required the Penman team ,NOT to imple,ment Halliday's concept of an integrated lexicogrammar, but to build instead a traditional lexicon which could be shared with the parser that was being contributed by another research team (working at Bolt, 7th Intemational Generation Workshop * Kennebunkport, Maine \u2022 June [21] [22] [23] [24] 1994 Beranek and Newman). There were some very interesting consequences of this decision.",
	"cite_spans": [
	{
	"start": 1211,
	"end": 1215,
	"text": "[21]",
	"ref_id": null
	},
	{
	"start": 1216,
	"end": 1220,
	"text": "[22]",
	"ref_id": null
	},
	{
	"start": 1221,
	"end": 1225,
	"text": "[23]",
	"ref_id": null
	},
	{
	"start": 1226,
	"end": 1230,
	"text": "[24]",
	"ref_id": null
	},
	{
	"start": 1231,
	"end": 1235,
	"text": "1994",
	"ref_id": null
	}
	],
	"ref_spans": [],
	"eq_spans": [],
	"section": "The Penman Upper Model",
	"sec_num": "4.4"
	},
	{
	"text": "First, recall that the choices in the system networks were at this time being increasingly seen as the 'meaning potential, and thus as semantic choices. In much of the detailed description in Halliday 1985, for example, the term semantic abounds. And the concept of 'realization' itselfthe process through which the choices in the system network are realized as st~ctures 7 is ~ne t, hat explicitly invokes the two levels of meaning and form. Thus if one does not represent the lexicon in system network form -as of course Penman does not -there is a major gap in the overall system network of the language at just tile place where one would wish to locate a system network of noun senses (to complement the list of noun forms). For/he same reason thePenman network lacks the meanings of all the other major word classes. This has made it harder than it is in the COMMUNAL framework (see below) to follow the increasingly strong trend in linguistics to give a more central place to lexis than it had in the 60s and 70s. (Fawcett Tucker and Lin 1993 show how lexis is handled in COMMUNAL, and important current modifications are exploiting yet further the value of having integrated lexicogrammatically realized networks of mean'ings -e.g. in verb complementation, and the many other cases where syntactic realizations depend upon lexical choice.)",
	"cite_spans": [
	{
	"start": 1020,
	"end": 1048,
	"text": "(Fawcett Tucker and Lin 1993",
	"ref_id": "BIBREF10"
	}
	],
	"ref_spans": [],
	"eq_spans": [],
	"section": "The Penman Upper Model",
	"sec_num": "4.4"
	},
	{
	"text": "The Penman response to the position in which they found themselves was a sensible compromise. The development work that might have gone into the meaning potential of lexical forms went partly into the separate lexicon (building into it the relevant features from the grammar), and partly into the Upper Model (UM). The main, function of the UMis to serve as an abstraction hierarchy (Bateman 1990:57) , and so to provide the usual functions of an ontology for property inheritance, etc. But at the same time,each concept in it is 'known to Penman', in the sense that it is possible to state for each concept in the UM the fragments of gran~matical or lexical realization that will be used to realize it (Bateman et al. 1990:5) .",
	"cite_spans": [
	{
	"start": 383,
	"end": 400,
	"text": "(Bateman 1990:57)",
	"ref_id": null
	},
	{
	"start": 703,
	"end": 726,
	"text": "(Bateman et al. 1990:5)",
	"ref_id": null
	}
	],
	"ref_spans": [],
	"eq_spans": [],
	"section": "The Penman Upper Model",
	"sec_num": "4.4"
	},
	{
	"text": "However, the section of the UM corresponding to noun senses seems surprisingly small for this task -at least, as described in Bateman et al 1990. It con, tains just over 30 categories (half being specialised within spatial-temporal'). The idea is that these are 'common core' concepts and that users of the UM will add to these as required.6 (It is interesting to compare these with the early semantic COMMUNAL categories, summarized in Section 5.)",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "The Penman Upper Model",
	"sec_num": "4.4"
	},
	{
	"text": "In Penman, then, the ontological categories map directly to (1)grammatically realizedoptions in the system network, and (2) the standard lexicon. The job does get done. But the great advantage of a SFG -namely that it can equally easily capture generalisations across large and small classes (including one-member classes) -is not exploited. Thus Penman fails to utilize the many advantages of a fully integrated semantic network that provides for direct andintegrated realizations in grammar and lexis (and indeed also in intonation or punctuataon).",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "The Penman Upper Model",
	"sec_num": "4.4"
	},
	{
	"text": "At the time when it was developed, in the very early 80s, Penman was a pioneering breakthrough. Its influence has changed the face of NLG. But the fact remains that it does not contain a lexicogrammar, in the SFG sense -just a grammar. The historic significance of Penman must not distract us from recognising that it has not yet explored Halliday s original proposal to integrate grammar and lexis in one great net-work. Yet this, as we in COMMUNAL have found, is a concept which -with the important modification that we do not restrict lexical meanings to the 'most delicate' parts of the network -brings enormous advantages of power and flexibility to the modelling of language.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "The Penman Upper Model",
	"sec_num": "4.4"
	},
	{
	"text": "If one's goals are (1) to build a SFG generator and (2) to relate it out to a representation of those ontological relation, s required for reasoning, the following question arises: Is it (a) necessary and (b) desirable to have two separate layers of network, representing (1) the semantic stratum within language and (2) the ontological relations in the belief system?'",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Alternative research strategies",
	"sec_num": "4.5"
	},
	{
	"text": "There are two research routes that may lead to a sound answer. The first is to try having just one layer of network, and then to move on to two it it turns out to be desirable or necessary. The second route to an answer is to start with two networks, one for each level, and then, if there turns out to be inadequate motivation for maintaining two, to abandon one or conflate/he two of/hem. In effect, Penman had the first strategy forced upon them, while we in COMMUNAl_, have followed the second. As we considered the purposes -and so the desirable structural characteristicsof the two potential components, it became increasingly clear that there are advantages in including them both.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Alternative research strategies",
	"sec_num": "4.5"
	},
	{
	"text": "The next two sections therefore set out the purposes and consequent structural organisation of (1) the system network for noun senses currently being implemented on a very large scale in COMMUNAL, and (2) the ontological aspects of the matchingpart of the belief system. While the discussion is naturally exemplified from the COMMUNAL Project, the principles are of general relevance to any researcher working m/his,area. H, ere we shall restrict ourselves to the core area of objects, including abstract and event-like objects, i.e. that part of the belief system that corresponds to the senses of nouns.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Alternative research strategies",
	"sec_num": "4.5"
	},
	{
	"text": "Thepurposes and structure of a system network for noun senses",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "5",
	"sec_num": null
	},
	{
	"text": "My work in linguistics and NLP over the last couple of decades has taught me that one of the most important lessons to learn, when trying to model language, is: DO NOT TRY TO DO TOO MUCH WORK AT ANY ONE LEVEL.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "The purposes of a system network for noun senses",
	"sec_num": "5.1"
	},
	{
	"text": "Thus the key to modelling language successfully is to have a sufficiently holistic theory and, to be able to recognize the appropriate level -or component -at which each particular type of work should be done.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "The purposes of a system network for noun senses",
	"sec_num": "5.1"
	},
	{
	"text": "Once one commits oneself to having a separate component to handle reasoning (includingproperty inheritance, etc) that is OUTSIDE LANGUAGE(even though, as I have always insisted, its internal structures are strongly INFLUENCED by language), then it becomes immediately clear that the system networks inside the language system are NOT in fact well-suited for use in reasoning. The reason for this is very simple: the design features that are required in the structure of-the relevant parts of the network are different from the design features required for ontologies.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "The purposes of a system network for noun senses",
	"sec_num": "5.1"
	},
	{
	"text": "So what are the purposes of/his part of the system network? Its primarypurpose is very simple. Just as the well known systems for transitivity, mood, and theme, etc generate clauses, so it is the purpose of this network to generate the nouns which will expound the heads of nominal groups.7 There is an important subsidiary purpose for those, noun senses to which participant roles (argument structure to some) are attached-, i.e. as with verb senses; compare the roles attached to die and death, to ascend and ascent, etc, but we cannot discuss these here. The other subsidiary purposes will be introduced in considering why the structure should have the form proposed below.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "The purposes of a system network for noun senses",
	"sec_num": "5.1"
	},
	{
	"text": "What, then, should the internal structure of a system network for generatino_ nouns be like? The answers given here are derived from t~e experience of developing very full sub-networks for large areas of the network for the 'cultural classification' of Ythings' in English -i.e. for the classification of objects provided by the noun senses of English. As an indication of the co, verage, ,consider the fairly well-developed sub-network for artefacts that are for human consumption' (i.e. food and drink): it has 137 systems and it generates 330 noun senses. Another quite well-developed area is that of plants, where there are 130 systems that generate 246 noun senses. (This reflects a layman's taxonomy of different trees and flowers; a specialist could well have three to s!x times as many.) An area currently under development, use of land, includes 'built up area', 'countryside, 'for_trav~lin-g', 'for_recreation',-'w~teland' and 'for_dividing_land, and it has so far 135 sub-systems, generating 370 noun senses. Many other areas are of similar size\u2022",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "The structure of a system network for noun senses",
	"sec_num": "5.2"
	},
	{
	"text": "We can give some idea of the overall taxonomic structure,by say!ng that 'use_of l,and' is one of fourteen types of artefact. The substantial (see below) features in the first few systems ,are as follows (where a system is represented as x -> y/z ): thing -> physical_thing / abstract thing / event thing; physical thing -> living thing / non living_thing;-living_thlng -~7 plant / creatu~; creature-> human_cr / non_human_cr; human_cr -> individual_hum / group hum; non living thing -> thing_as object/general _physTcal_phenoTnenon7 thing_as_substanCe; thing_as object -> artefact / natural object; general physical phenomenon r> energy / weatffer / colour; eve~_thing -5 event _as_process / complex_event. The network is already very large, and it is growing all the time.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "The structure of a system network for noun senses",
	"sec_num": "5.2"
	},
	{
	"text": "What, then, are the pri,nciples on which it is constructed? The simplest 'network for the noun senses of English would consist of just one large system, containing one long list of the noun senses, each realized by a noun form -i.e. a single, massive system. Why not do this? The main linguistic evidence is the existence of hyponymic and contrastive relations between words -or more strictly, between word senses. Thus the word-form cat is a hyponym of mammal, because the ,meaning, 'cat' is systemically dependent on the meaning mammal'. And a system network is equally appropriate for giving formal expression to another major type of relationship between words, e.g. as specified in stand~d works on semantics such as Lyons 1977: that of 'contrast; the sense of dog is in contrast with that of cat The evidence that such hypon~/mic and contrastive relations are anc,hored ,firmly within language -and not simply in some higher component of the behef system, is the existence at the level oi ~ form of nouns such as thing, object, animal, mammal, person, etc; see the note on the as such type of feature in 5.5.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "The structure of a system network for noun senses",
	"sec_num": "5.2"
	},
	{
	"text": "The overall structure is therefore taxonomic. Within",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "The structure of a system network for noun senses",
	"sec_num": "5.2"
	},
	{
	"text": "\u2022 tcr \u2022 \u2022 i \u2022 \u2022",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "The structure of a system network for noun senses",
	"sec_num": "5.2"
	},
	{
	"text": "this framework, we turn now to a ~mdehne cntermn that derives directly from the practical experience of building ontologies.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "The structure of a system network for noun senses",
	"sec_num": "5.2"
	},
	{
	"text": "Our experience in COMMUNAL suggests that it is advisable to avoid s, imultaneous entry conditions, i.e. rightopening curly 'and brackets. Here Dahlgren's point about overgeneration (4.3 above) DOES apply. (We shall meet the formalism when we consider ontological relations in Section 6, because in that type of network they do have a role to play.) Theproblem in generation is that such parallel systems lead to parallel pathways through the network, and so, almost inevitably, to permitting more coselections than are in f,act possible. For example, the distinction between 'male and'female' is realized lexically in the case of only some animals. In COMMUNAL we simply allow the repetition of such systems, or gather the relevant features together into a disjunctive entry condition (as in Figure 2 ), whichever is easier in practical terms for the maintenance of the network as it is expanded. The key point is that, if a network designed to generate nouns allows one to go downpathways -and so to choose a set of features -FOR WHICH THERE IS NO REALIZATION, it is a bad system network.9 (But note that, while simultaneity tends to lead to problems in networks for lexis, it has a valuable role in modelling g(ammatically-realized meaningsand, as we shall see in Section 6, ontological relations.)",
	"cite_spans": [],
	"ref_spans": [
	{
	"start": 792,
	"end": 800,
	"text": "Figure 2",
	"ref_id": null
	}
	],
	"eq_spans": [],
	"section": "The structure of a system network for noun senses",
	"sec_num": "5.2"
	},
	{
	"text": "The large network for noun senses in COMMUNAL therefore has essentially the structure illustra,ted, in Figure 1 . Neutrally, this structure can be described as: If a' then 'b' or 'c'., ,M. ore typically, in a SFG framework, we express it as: 'If a is chosen (by a speaker or a computer generator) then either 'b' or 'c' must be chosen.' Each feature may become the entry condition to a subsequent system, thus building up a ,s~,stem,network, so it might cog, tl,nue: 'If 'b' is chosen, then d' or 'e must be chosen, and if c is chosen then f or 'g' or 'h' must be chosen .... ' and so on.",
	"cite_spans": [],
	"ref_spans": [
	{
	"start": 103,
	"end": 111,
	"text": "Figure 1",
	"ref_id": null
	}
	],
	"eq_spans": [],
	"section": "The structure of a system network for noun senses",
	"sec_num": "5.2"
	},
	{
	"text": "Figure 1\" A simple system network However, there is one type of complex entry condition which we have found to be occasionally useful in the network for noun senses: the disjunctive entry condition. This has the form shown in Figure 2 Here is a (simplified) example of this phenomen,,on,,from the current network (where \"/\" signals or' and -> signals 'entry to a system'):",
	"cite_spans": [],
	"ref_spans": [
	{
	"start": 226,
	"end": 234,
	"text": "Figure 2",
	"ref_id": null
	}
	],
	"eq_spans": [],
	"section": "Ii f I:",
	"sec_num": null
	},
	{
	"text": "tomato_plant_c / pepper plant_c / cucumber plant c / strawberry_plant_c -> 95% plantness_explicit (66.325) 5%",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Ii f I:",
	"sec_num": null
	},
	{
	"text": "plantness_implicit.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Ii f I:",
	"sec_num": null
	},
	{
	"text": "This representation of a system network (in the COMMUNAL format) is to be read as follows:",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Ii f I:",
	"sec_num": null
	},
	{
	"text": "'If you select either 'tomato plant', etc (whic,h come from the ve etable sectmn of the ~,round crops network)or",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Ii f I:",
	"sec_num": null
	},
	{
	"text": ", d~ v \u2022 -- o \u2022 v",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Ii f I:",
	"sec_num": null
	},
	{
	"text": "strawberry_plant (which comes from the fruit subnet,work), then there is a choice between having the concept of plantness' made specific and leaving it implicit.' So, if the choice is to make it explicit, forms such as tomatoplant will be generated, an, d, if implicit, forms such as tomato -but in the gardeners sense of 'tomato plant, as in l've watered the tomatoes this morning.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Ii f I:",
	"sec_num": null
	},
	{
	"text": "Another valuable characteristic of the COMMUNAL system networks is the use of probabilities. As ,you can see, the weighting is 95% to 5% towards explicit (unless this is overruled by the discourse planner). The reason for this strong weighting is that there is another type of object 7th International Generation Workshop \u2022 Kennebunkport, Maine \u00b0 June [21] [22] [23] [24] 1994 that is commonly referred to by the same word-form, namely the fruit of the tomato plant. (The choice of 'plantness explicit' triggers Realisation Rule 66.325, which adds to the head of the nominal group the item plant. )",
	"cite_spans": [
	{
	"start": 352,
	"end": 356,
	"text": "[21]",
	"ref_id": null
	},
	{
	"start": 357,
	"end": 361,
	"text": "[22]",
	"ref_id": null
	},
	{
	"start": 362,
	"end": 366,
	"text": "[23]",
	"ref_id": null
	},
	{
	"start": 367,
	"end": 371,
	"text": "[24]",
	"ref_id": null
	},
	{
	"start": 372,
	"end": 376,
	"text": "1994",
	"ref_id": null
	}
	],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Ii f I:",
	"sec_num": null
	},
	{
	"text": "A third criterion used in constructing the network relates to its value in specifying what in Chomskyan linguistics are termed 'selectiona] restrictions'. Some might wish to argue that these really belong in the belief system rather than in language itself. However, it turns out that it is quite simple, if one constructs the network on appropriate criteria, to enable very many of the types of restriction that it was hoped to capture in transformational models of the 1960s and 1970s to be captured in a system functional grammar. Moreover, since we use probabilities in the grammar in any case (for purposes that we cannot go into here), we can ,capture selection restrictions as relatively strong or weak preferences'. Thus there seems to be no reason why we should not c.apture such phenomena in both the system network within language and in the belief system. If we can do this, it brings the advantage, when TESTING the system, that the lanzua~e comoonents can be un independently of the behef system. They can be set to generate sentences randomly as we test the lexicogrammar, and still generate sentences that sound fairly plausible.10",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Ii f I:",
	"sec_num": null
	},
	{
	"text": "The way this works is, in outline form, as follows. Suppose we have just generated a clause, with an Agent as Subject and ask as the main verb. We want to restrict the choices when the network is re-entered, so that, if a nominal gAroup with a noun at its head is to be genemt~ to fill the gent, it will have one that is plausible as an asker. We do this by specifying that, on re-entry to the network, the following features are chosen: [thing, concrete, living, creature, 99.9% human / 0.1% non_human, whole_hum, 99% individual_hum / 1% group_hum]. ,At two points, you will notice, there is not an absolute preselection' of just one of the features in the system, but a preference for one over another. Thus the feature [human] Is shown to be a thousand times more probable than [non_human] -thus allowing for talking computers and talking animals, such as the white rabbit in Alice in Wonderland. Rather similarly, individual humans are shown as being a thousand times more likely to be the Agents in a process of 'asking' than are groups of humans -though groups, such as a committee, may well ask questions on occasions.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Ii f I:",
	"sec_num": null
	},
	{
	"text": "We turn now to an important issue that arises as a result of taking this position, and we shall illustrate it first from English.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Ii f I:",
	"sec_num": null
	},
	{
	"text": "One important effect of deciding to organize the network around meanings (rather than around grammatical correlates ,of wor, d forms) is to put in perspective the contrast between \u2022 count and 'mass' noun senses that is so dominant in English (and many other European languages). If, for example, we are stating preferences for the Affected entity in a process of 'eating', it is not important whether what is eaten is mass or count. The COMMUNAL network for food classifies types of food on semantic criteria -so that, for example, vegetables that typically occur with ,a meat course are placed together, with the mass noun sense cabbage' next to the count noun senses of 'potato' or 'pea'. Moreover, the network also includes a way of showing that, while potato' occurs regularly as either singular or plural, it is rather unusual, to talk of a single pea. The result is that the system, knows that it is a tho, usand times more likely that \"peas will be generated than pea'. (Note that the system networks in COMMUNAL not only have probabilities, but the grammar can chan,~e these when required. For a more detailed account of t~is, together with a full worked example, see Fawcett, Tucker and Lin 1993.) raowever, there is not in fact a neat 'count' versus 'mass' distinction in English at all. It is one of those useful gener-alisations which hold for 95% of the time, but which, if one commits oneself to it, leads to considerable trouble when the idea is extended to the whole of langua,~e. It is certainly not a category that can be extended, on thgbasis of a system network for English, even to a close European !anguag,e such as French. The illogicality -in physical number terms -of items such as furniture and cutlery is just the well-known visible tip of quite a large iceberg. First, note the 'plural-only' nouns, such as police, staff and contents. Theft there are the 'plural-prefern'ng' items,--with varying strengths of preference, as for example betwee,n pea ana sprout, and pebble and leaf. There are also the paironly' nouns such as trousers, scissors, and binoculars -and, even though the word denotes two garments, of which only the bottom half conforms to the pattern, there is pyjamas.",
	"cite_spans": [
	{
	"start": 1176,
	"end": 1206,
	"text": "Fawcett, Tucker and Lin 1993.)",
	"ref_id": "BIBREF10"
	}
	],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Problem case 1: the 'count' versus 'mass' contrast",
	"sec_num": "5.3"
	},
	{
	"text": "As an example of the dissonance within one relativel3; smal! semantic field between the grammatical criterion of 'number and the semantic classification of noun senses, consider the field of clothing. Suppose the problem is that of stating the preferences for the entity that is to complete a clause such as He went home and put on ..... It could be a nominal group with, at its head, (1) a mass noun such as some warm clothing, or (2) a plural-only noun as in some warm clothes, or (3) a singular noun such as a warm jersey, or (4) a pair-only noun such as some warm trousers.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Problem case 1: the 'count' versus 'mass' contrast",
	"sec_num": "5.3"
	},
	{
	"text": "The COMMUNAL solution to this problem is as follows. Every feature in the syste, m, network f o, r which there is a realization is given a suffix c (for 'count thinzs) or ~_m' (for 'mass' things,) or '_pl' (fo~ plural only' thinks) or _pair' (for pair only' things). Those labelled' c' lead into the system for NUMBER, where there is achoice between [singular] and [plural], while for the others there is no choice: they are either [mass] or [plural] . Those for which a RELATIVE preference for [plural] has been stated will enter a version oT the NUMBER system for which the probabilities have been re-set according to the strength of the preference associated with than noun sense, i.e. the lexicogrammar ,'knows' that for 'pea' the probabilities of choosing 'plural are very much greater than for 'cabbaze'.",
	"cite_spans": [
	{
	"start": 432,
	"end": 438,
	"text": "[mass]",
	"ref_id": null
	},
	{
	"start": 442,
	"end": 450,
	"text": "[plural]",
	"ref_id": null
	},
	{
	"start": 495,
	"end": 503,
	"text": "[plural]",
	"ref_id": null
	}
	],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Problem case 1: the 'count' versus 'mass' contrast",
	"sec_num": "5.3"
	},
	{
	"text": "Consider too the question of where to place th~ two senses -realized as count and mass nouns -of cloud. In this case the differences between the two appear to derive mainly from the fact that one is an individual entity and the other is a non-discrete mass of 'stuff; they are not differentiated by function. So the system network shows them to be distinguished as follows: cloud-> cloud as individual/cloud as mass.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Problem case 1: the 'count' versus 'mass' contrast",
	"sec_num": "5.3"
	},
	{
	"text": "Notice that this is a rather different matter from 'oak' as 'individual' and as 'mass'; in the first case the referent is a tree (with oak-tree as a possible variant, if 'treeness explicit' is selected), and it is located with the rest of the tr-eegas a",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Problem case 1: the 'count' versus 'mass' contrast",
	"sec_num": "5.3"
	},
	{
	"text": "\u2022uP e of plant, while in its other sense it is a material whose nction is to be used for making things, and it is located with other types of wood, fairly c\u2022ose to 'iron', etc.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Problem case 1: the 'count' versus 'mass' contrast",
	"sec_num": "5.3"
	},
	{
	"text": "In the COMMUNAL system it is even possible to build in the preference for meanings realized in expressions such as a pair, followed by of, as in a pair of trousers. This is because trousers rather thanpair is treated as the head of the nominal group. The words a pair are generated as a special nominal group expressing 'quantity' that fills the quantifying determiner. It is because the noun sense of the nominal .qroup is generated on the first pass through the network ~or the object i.n question -I.e. because the lexically realized meaning is integrated with the grammatically realized meaning in C ONIMUNAL-that part of the realization of the choice of trousers' can be to express a preference for the way in which the auantifvin,~ determiner is filled -here by the embedded nominal grou'p o-f a parr.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Problem case 1: the 'count' versus 'mass' contrast",
	"sec_num": "5.3"
	},
	{
	"text": "Thus the COMMUNAL way of handlinz these various aspects of 'number' in English can reflect-accurately the 7th International Generation Workshop \u00b0 Kennebunkport, Maine \u2022 June [21] [22] [23] [24] 1994 true, complex nature of 'number' and 'quantity' in English, while at the same time maintaining the semantic relationships of hyl~onymy and contrast in the network, and so making possible the expression of preferences.",
	"cite_spans": [
	{
	"start": 174,
	"end": 178,
	"text": "[21]",
	"ref_id": null
	},
	{
	"start": 179,
	"end": 183,
	"text": "[22]",
	"ref_id": null
	},
	{
	"start": 184,
	"end": 188,
	"text": "[23]",
	"ref_id": null
	},
	{
	"start": 189,
	"end": 193,
	"text": "[24]",
	"ref_id": null
	},
	{
	"start": 194,
	"end": 198,
	"text": "1994",
	"ref_id": null
	}
	],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Problem case 1: the 'count' versus 'mass' contrast",
	"sec_num": "5.3"
	},
	{
	"text": "The issue is in fact much wider than that of whether 'count' vs. 'mass' should be a primary distinction in English and related !anguages. How strong a candidate it is for a generalized interlingua' ontology that is to accommodate all the languages of the world? Consider Chinese, with its well-known classifier system, in which the mass-count distinction plays no part. Then think about Swahili, with its ki-viclass of non-living things, its m-waclass for humans, its u-class for abstract things, etc. Japanese, as it happens, has a special set of cardinal determiners, whose form depends on the semantic class of the noun: i.e. whether the object is human or a small thing or even, it would seem, a long thin thing.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Problem case 2: long thin things and other such grammatically realized categories",
	"sec_num": "5.4"
	},
	{
	"text": "Thus, \"if the thing concerned is a flower (hana), a tree (ki), a pen (pen), a pencil (enpitsu) or a,rive, r (kawa)all longthin things -the determiner meaning one will be ippon. But if the thing is a human it is hitori, and if a non-human creature it is ipipiand so on, for many more classes of thing and for many more cardinal determiners. The semantic generalisation that unites those things that require ippon appears to be simply that they are all \"long thin things'. If this seems strange to the investigator from a European background, consider how odd it must seem to the investigator of English from outside Europe who finds that, in English and related languages, there exists a basic distinction which affects many aspects of the semantics and syntax of the nominal group, between things that are and things that are not 'countable'.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Problem case 2: long thin things and other such grammatically realized categories",
	"sec_num": "5.4"
	},
	{
	"text": "However, it must be emphasised again that, important as the distinction is at many points, it is not the basic organising p.rinciple of the semantics of English noun senses. While there is, of course, a distinction between 'substances' and 'objects' in the taxonomy implied in the COMMUNAL network, and while all substances are 'mass', it is not the case that all 'objects' are 'count' -as we saw in section 5.3.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Problem case 2: long thin things and other such grammatically realized categories",
	"sec_num": "5.4"
	},
	{
	"text": "In COMMUNAL, then, our semantic system network gives no weight to grammatically realized contrasts such as count' versus 'mass', and we use instead the semantic criteria such as those that help us to state the preferences associated with a given participant role such as agent.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Problem case 2: long thin things and other such grammatically realized categories",
	"sec_num": "5.4"
	},
	{
	"text": "Is there a price to be paid for all of the advantages outlined in the, preceding sections? The answer seems to be that there isn t. If the mechanism for handling NUMBER in a way that is dependent on noun senses (in 5.3) required one to work from a visual representation of the wiring, it would be tedious in the extreme But in the computer it ~s a simple matter -and this has in turn suggested a simple representation for the written version.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Problem case 2: long thin things and other such grammatically realized categories",
	"sec_num": "5.4"
	},
	{
	"text": "In a full system network for nou, n senses, the relations between features are not all of the subcategorization type that the systemic notation typically signifies. The extensive work that has been done in COMMUNAL over the ,last few years has produced a small set of supplementary (or special ),features whose function is to express relations between the substantial' features that represent noun senses. In principle, all of those shown in Figure 3 can occur between any two substantive features. Here, then, 'x' stands for a feature such as 'human', and 'a', 'b' and 'c' are the next 'substantive' features. Possible realizations of the selection of the feature (in some cases in features dependent on substantive features) are given in Figure 3 . Figure 3: Some types of features that are not related by 'betype'relations 'Substantive' features are those that translate types of object in the belief system. In most cases none, or only one, of the special features are needed bet w, een substantive features. Note in particular the 'xas-such feature. This provides for the generation of those nouns that are used for Iexical substituUon such as thing, stuff and, more delicately, plant and animal. It is the existence of such forms that provides the evidence that such less delicate meanings exist in natural languages, and that the intra-linguistic level of semantics does indeed require a network for it to be adequately represented.",
	"cite_spans": [],
	"ref_spans": [
	{
	"start": 442,
	"end": 450,
	"text": "Figure 3",
	"ref_id": null
	},
	{
	"start": 740,
	"end": 748,
	"text": "Figure 3",
	"ref_id": null
	}
	],
	"eq_spans": [],
	"section": "'Special features' in the system network",
	"sec_num": "5.5"
	},
	{
	"text": "There are possible variations of the feature [xspecified] which it is important to note. The type of specification may be spelled out more fully, with a feature corresponding to each type that leads to its own dependent network. Thus in some cases (such as [human]) we may find Ix_as_such] vs. Ix_specified_by_form] vs. Ix_specified by_role].",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "'Special features' in the system network",
	"sec_num": "5.5"
	},
	{
	"text": "The somewhat dense description given above is intended to give the flavour of the criteria that are guiding the construction of the very large system network for noun senses in COMMUNAL.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "'Special features' in the system network",
	"sec_num": "5.5"
	},
	{
	"text": "There is one last benefit that this approach brings. It is a practical rather than a theoretical one. This is that the process of constructing the network, and so of deciding which types of 'special feature' are needed, goes some way in preparing the ground for constructing the equivalent ontological aspects of the belief system. And it is this to which we now turn.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "'Special features' in the system network",
	"sec_num": "5.5"
	},
	{
	"text": "6 The purposes and structure of the ontological aspects of a belief system",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "'Special features' in the system network",
	"sec_num": "5.5"
	},
	{
	"text": "What are ontologies for? They are as they are, of course, because of the functions that they are required to perform. In the case of an ontology, the purpose is NOT to represent the meanings of the nouns of a language, but to facilitate reasoning.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "The purposes of an ontology",
	"sec_num": "6.1"
	},
	{
	"text": "Thus lexically prominent register distinctions such as that between fag and cigarette would be modelled as de,noting t,he same generic object, and so share the same concept. This is because the same events Or 'propositions') are attached to it, whatever degree of rmality is selected in the TENOR system. It is the latter that requires recognition in the belief system, and not a difference of concept.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "The purposes of an ontology",
	"sec_num": "6.1"
	},
	{
	"text": "Ontologies are used in two principal ways, the second beinz deoendent on the first. The first is for the type of reaso-nin~known as entailment (see, for example, Leech 1969 and 1974/81) . Through entailment a reasoner can infer from the belief (or proposition) that Object 79 is a member of the set, of objects denoted by the form dog (and by the sense 'dog) that it is also a member of the set of objects to which the sense 'mammal' pertains. In layman's lan_oua2e, A dog is a mammal. The crucial point Is that, typl~caffy, the directionality of the reasoning is from the more delicate category to the less delicate. For examp!e, in terms of Figure 4 below, if an object is 'd' it is also b' 7th International Generation Workshop \u2022 Kennebunkport, Maine \u00b0 Jmae [21] [22] [23] [24] 1994 and 'a'. This directionality is of course different from the typical use of a system network in generation.",
	"cite_spans": [
	{
	"start": 162,
	"end": 176,
	"text": "Leech 1969 and",
	"ref_id": "BIBREF16"
	},
	{
	"start": 177,
	"end": 185,
	"text": "1974/81)",
	"ref_id": null
	},
	{
	"start": 761,
	"end": 765,
	"text": "[21]",
	"ref_id": null
	},
	{
	"start": 766,
	"end": 770,
	"text": "[22]",
	"ref_id": null
	},
	{
	"start": 771,
	"end": 775,
	"text": "[23]",
	"ref_id": null
	},
	{
	"start": 776,
	"end": 780,
	"text": "[24]",
	"ref_id": null
	},
	{
	"start": 781,
	"end": 785,
	"text": "1994",
	"ref_id": null
	}
	],
	"ref_spans": [
	{
	"start": 643,
	"end": 651,
	"text": "Figure 4",
	"ref_id": null
	}
	],
	"eq_spans": [],
	"section": "The purposes of an ontology",
	"sec_num": "6.1"
	},
	{
	"text": "The second main use of ontologies is an extension of this. It is for the inheritance of what are commonly called 'properties'. (However, as we shall see, the term 'oroperty' is potentially misleading).",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "The purposes of an ontology",
	"sec_num": "6.1"
	},
	{
	"text": "Thus, if such 'iarop'erti'es' are attached to a category at one node of the ontology, they can then be assumed to hold for any given object to which a logically dependent category applies., So, if we build into a belief system the proposition that land mammals typically have four limbs', then we can infer that, because a dog is a land mammal, it too typically has four limbs -and so on. (More strictly, as we shall shortly see, the relationship is between the generic thing that corresponds to that node; in the COMMUNAL model of logic dogs is a referring expression as much as is this dog.)",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "The purposes of an ontology",
	"sec_num": "6.1"
	},
	{
	"text": "These types of entailment-based reasoning are important. But it is equally important to recognize that they are only one of a variety of types of reasoning that are regularly used in real life situations. Moreover, the concept that all such beliefs about categories of object are handled by inheritance is not a matter that is entirely beyond dlspute; It is arguable, for example, that the category of human is so prominent in our perception of the world that we build sets of beliefs around it -andthat we do not use an ontolo,qical structure in order to inherit, every time that we refer tGa human, all of the many propositions that relate to the many other, ever more general, categories that are superordmate to 'human' -such as 'mammal', 'creature', 'living' and 'concrete'. And, once one allows for this possil~ility, -there is no clear way of knowing where to stop. There may be many such nodes that laave attached to them large sets of beliefs (or 'propositions ) that are held by the system, and which may indeed involve the redundant repetition of beliefs that are attached to less delicate concepts. Who can say whether it is more economical (or more ele_~ant more efficient, or whatever) to store a large (but not l~fir~ite) set of propositions many times over at tile various nodes where they are most often needed -the basic types', in the terms of Rosch (1978) -or to store each of them just once but to have to perform a multiple act of entailment reasoning, involving multiple searches back down the tree, every time one u,ses the belief that a dog needs air to breath, or is a 'creature, or is a 'concrete' object? A particular dog-lover, for example, may have his/her set of primary beliefs about spaniels attached to 'spaniel', rather than to 'dog', and so on (cp. Reiter and Dale 1992) . In other words, while we are undoubtedly capable of performing the quite complex type of reasoning involved in inheritance, it may be that it is not the backbone of all reasoning that it has sometimes been assumed to be.",
	"cite_spans": [
	{
	"start": 1363,
	"end": 1375,
	"text": "Rosch (1978)",
	"ref_id": "BIBREF26"
	},
	{
	"start": 1785,
	"end": 1806,
	"text": "Reiter and Dale 1992)",
	"ref_id": "BIBREF25"
	}
	],
	"ref_spans": [],
	"eq_spans": [],
	"section": "The purposes of an ontology",
	"sec_num": "6.1"
	},
	{
	"text": "The ar_oument is not that inheritance has no place in our reasoning,'but (1) that it may have a much less central place than has generally been supposed, and (2) that other types of reasoning are probably equally or mo, re important. Abelief system of the type assumed here is object-oriented in the sense that it consists of a vast number of specific objects and generic objects (with each of the former linked by a 'be-an-instance-of relationship to one or more of the latter) such that one of the many things that the system believes about any generic object is what other generic object -or objects -that generic object is itself a type of.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "The purposes of an ontology",
	"sec_num": "6.1"
	},
	{
	"text": "In the above discussion, we have been assuming that we are considerin,~ generic objects, e.g. cats in general. I assume that t~ere would be general agreement that the relationship between 'mammal' and 'cat' is what we might term a 'be-a-type-of relationship, while that between a specific instance of a cat, such as our family cat Timmy, is a 'be-an-instance-of relationship. Specific objects are related to concepts via the belief that 'Timmy is an instance of a cat'. In other words, we need to distinguish between (1) 'Timmy is a cat' and (2) 'Cats are mammals.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "The purposes of an ontology",
	"sec_num": "6.1"
	},
	{
	"text": "I suggested earlier that the use of the term 'properties' to refer to the 'propositions' attached to categones can be misleading. Consider the simplest mod, el of ,what 'properties' shouldbe attached to categories (the frame approach). All there is space to say here is that there are clearly limitations as to what can be expressed in such limited structures, and, like others, we in COMMUNAL are exploring richer alternatives. In our case we are experimenting with a specially developed logical form for the representation of both (I) complex beliefs about, say, the mating habits of dogs, and (2) the belief that dogs are mammals -and indeed the belief that dogs are typically pets ('multiple inheritance').",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "The purposes of an ontology",
	"sec_num": "6.1"
	},
	{
	"text": "The purpose of ontological relations, then, is to facilitate reasoning. But our prediction is that in future there will be less emphasis on inheritance and more on other logical relationships. Given these purposes, we turn now to the question of the structure of ontologies.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "The purposes of an ontology",
	"sec_num": "6.1"
	},
	{
	"text": "There seems to be a general agreement that an ontology has the general form of a taxonomy' or 'hierarchy' (in one sense of that overused term) or 'tree (which is equal!.y overused). The type of 'tree' required here is thus a paradigmatic tree (rather as is a system network) where, in the simplest model, a pathway that consists of a list of features chosen as one traverses the network corresponds to any one (as p,ect of) an object. In t h,e simplest type of taxonomy, then, a is subcategofized as b or 'c ~, 'b' as' d' or 'e', 'c' as 'f, 'g' or ' h', and so on; see Figure 4 . d a~ b~ef",
	"cite_spans": [
	{
	"start": 505,
	"end": 549,
	"text": "'c ~, 'b' as' d' or 'e', 'c' as 'f, 'g' or '",
	"ref_id": null
	}
	],
	"ref_spans": [
	{
	"start": 569,
	"end": 577,
	"text": "Figure 4",
	"ref_id": null
	}
	],
	"eq_spans": [],
	"section": "The structure of ontologies",
	"sec_num": "6.2"
	},
	{
	"text": ".h Figure 4 : Relations in the simplest type of ontology Such a network show, s that, if an object satisfies the conditions for being an h ,it follows that it is also a 'c', and so also an a'. So far, this structure is similar to the simplest type of system network, as ,shown in Figure 1 . This simplest structure for an ontology is in practice au,~mented in various ways in all of the ontologies that I kn~w of. The essential addition is as in Figure 5 : This is to be read as, 'a' is subcategorized as 'b ', 'c' or 'd', and also as 'e' or 'f; 'b' is s, ubcategonzed as 'g' or 'h', and 'd' as T, 'j' or 'k' , and also as i or m.",
	"cite_spans": [
	{
	"start": 508,
	"end": 608,
	"text": "', 'c' or 'd', and also as 'e' or 'f; 'b' is s, ubcategonzed as 'g' or 'h', and 'd' as T, 'j' or 'k'",
	"ref_id": null
	}
	],
	"ref_spans": [
	{
	"start": 3,
	"end": 11,
	"text": "Figure 4",
	"ref_id": null
	},
	{
	"start": 280,
	"end": 288,
	"text": "Figure 1",
	"ref_id": null
	},
	{
	"start": 446,
	"end": 454,
	"text": "Figure 5",
	"ref_id": "FIGREF1"
	}
	],
	"eq_spans": [],
	"section": "The structure of ontologies",
	"sec_num": "6.2"
	},
	{
	"text": "Here the right-opening curly brackets do NOT signify that you must follow all the designated paths, as they would in a system network -because this type of network is desig.ned to be used from right to left, for the types of entadment and inheritance outlined above. Dahlgren (1988:46f.) discusses the mathematical properties of ontologies, and she quite rightly points out that, while binary ontologies 'have simplifying mathematical properties, they are not likely to be .psychologically reap (e.g. suggesting that we operate with 'fish ' vs. 'bird' vs. 'mammal', etc) . She also points out the need for cross-classification, in that we classify animals, let us say, both in terms of their 'types' (a concept that itself needs to be unpacked, at least for some purposes), but a!so in terms of their 'role' in humankind s scheme of things ( tame' vs. 'wild' animals, probably with other such subcategories as well)\u2022 Thus Dahlgren and others operate with the formalism shown in Figure 5 . In the cases of other ontologies there is even more freedom, in that any feature can be linked by a line that represents a 'be-type' relationship to any feature to its left. And in some ontolbgies, of course, other relationships such as 'be-a-part-of' are used.",
	"cite_spans": [
	{
	"start": 267,
	"end": 287,
	"text": "Dahlgren (1988:46f.)",
	"ref_id": null
	},
	{
	"start": 539,
	"end": 570,
	"text": "' vs. 'bird' vs. 'mammal', etc)",
	"ref_id": null
	}
	],
	"ref_spans": [
	{
	"start": 978,
	"end": 986,
	"text": "Figure 5",
	"ref_id": "FIGREF1"
	}
	],
	"eq_spans": [],
	"section": "The structure of ontologies",
	"sec_num": "6.2"
	},
	{
	"text": "However, there is an important sense in which the above diagrams are misleading -at least in relation to the current COMMUNAL framework. This is because we consider that the relations do not hold, in fact, between the 'concepts' in the ontolo~ies, but between generic obiects. The relationship that we\"have shown in Figures 4 and5 by a line can also be expressed by the following: e (el 11, [ca:o222, pr:bejype, at:o333] all do=s, while object no. 333 is defined as mammals. There are several aspects of the meaning of this representation of a belief which are assigned by default -,most importantly, that the time position of the event is past, present and future', that the' 'confidence level' is so high as to be interpretable as 100% confident'. Taken together, they state, m a natural language translation: All dogs are mammals. This event-based representation is used in order to enable the system to carry out reasoning on inputs to and outputs from the system that have the sorts of annoyingly messy complexity associated with natural languagerepresentations of time position, usuality and quantification, for example\u2022 In the belief systems of the future it will no longer be possible, in our view, to depend on simple data structures such as frames. The need to incorporate more sophisticated representations of time, of modality and of other such phenomena related to events demands that infor,mation be stored in the form of some type of 'predicate logic, e.g. as in the (simplified) example above\u2022",
	"cite_spans": [
	{
	"start": 383,
	"end": 420,
	"text": "(el 11, [ca:o222, pr:bejype, at:o333]",
	"ref_id": null
	}
	],
	"ref_spans": [
	{
	"start": 316,
	"end": 330,
	"text": "Figures 4 and5",
	"ref_id": "FIGREF1"
	}
	],
	"eq_spans": [],
	"section": "The structure of ontologies",
	"sec_num": "6.2"
	},
	{
	"text": "The key point is that the representation, like the minimal operational syntactic unit otthe clause, is based on the concept of an EVENT (our equivalent term to 'proposition' and 'eventuality' in other frameworks). It uses categories that reflect idealised aspects of systemic functional grammar, so it is a 'systemic functional logical form' (SFLF).",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "The structure of ontologies",
	"sec_num": "6.2"
	},
	{
	"text": "Note, finally, that the relationship that it expresses is NOT one that holds between two concepts, but between two referring expressions, each with its SFLF structure, and each of which refers to a generic object\u2022 To adopt this position leads to a reappraisal of the status of dmgrams representing ontological relationships such as those in Figures 4 and 5. Those relations are, m the approach advocated here, simply one event type among many within the mass of belief, s that the system holds about dogs\u2022 In other words, the fact' (i.e. the confidently held belief) that dogs are mammals is just one of many things that the system assumes that it 'knows' about the generic object of 'dogs.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "The structure of ontologies",
	"sec_num": "6.2"
	},
	{
	"text": "The COMMUNAL Project (Fawcett, Tucker and Lin 1993) has demonstrated the immense advantages that follow from having a unified system network for all meanings, whether reafized grammatically or lexically. Thus Hallida~'s original 1961 insight was well-founded, the only major modification needed being that lexically realized meanings are not necessarily most delicate', in the sense of 'at or near the terminal leaves of the system network. (Meanings realized in intonation should similarly be integrated into the overall network, and the way in which thls is done in COMMUNAL is described in Fawcett 1980.) In other approaches some equivalent intralinguistic level of 'linguistic meaning' would surely be of similar value, It would be interesting to compare the COMMUNAL 'division of labour in modelling lexical meaning with other current proposals, such as those of Reiter and Pustejovsky, but that is beyond the scope of this paper. What is clear is that information about the purposes oil entities, how they come into being, etc is all handled through events attached to the equivalent genetic object in the belief system\u2022 The two types of network -the system network of noun senses and the ontological relations in the belief systemserve different purposes. One is to generate nouns (and to control relateOstructures such as those associated with participant roles) and the other is to facilitate reasoning. In the first it isprefer,able to avoid right-opening brackets -but in the second such cross-classifications' (or their equivalents \u2022 I ' 0\" \" \" m an event-based lo=lcal form) are wtal to certain types of reasoning\u2022 Since the two structures serve different functions, both have their place in a holistic model of language. (I say 'structures, because it is probably unhelpful, in the new framework, to think of the e, quivalen, ts of ontological relations in the belief system as a network .)",
	"cite_spans": [
	{
	"start": 21,
	"end": 51,
	"text": "(Fawcett, Tucker and Lin 1993)",
	"ref_id": "BIBREF10"
	},
	{
	"start": 593,
	"end": 607,
	"text": "Fawcett 1980.)",
	"ref_id": "BIBREF6"
	}
	],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Summary and conclusions",
	"sec_num": "7"
	},
	{
	"text": "In the previous section we have seen how the COMMUNAL Project has begun to develop a new way of representing -and so thinking about -the relations found in standard ontologies. Since these are now seen as a part of the belief system, the suggestion !s that we no longer need to think in terms of 'an ontology, in the sense of a taxonomy of 'concepts' -but rather in terms of relations between genetic objects. We think that these fit more naturally into the more complex, event-based reasoning that we judge will be needed for the systems of the future\u2022 So perhaps there is no such thing as an ontology, but instead a set otmutually referrin,2 beliefs, each represented as an event?",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Summary and conclusions",
	"sec_num": "7"
	},
	{
	"text": "If tKis approach is on the right lines, or even if it is only partly on the ri,2ht lines, it is time to be moving on to the exploration of t~e next generation of belief systems. These will make different assumptions about the nature of the relationship of word forms to word senses, and word senses to the related aspects of the belie, f system -and perhaps too about the nature of 'ontological relations.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Summary and conclusions",
	"sec_num": "7"
	},
	{
	"text": "! The research reported here was supported by grants from DRA Malvem (under contract no. ER1/9/4/2181/23), from the University Research Council of International Computers Ltd, and from Longman, and by the University of Wales, Cardiff. I am grateful to my colleagues Gordon Tucker, Yuen Lin and Ulrich Gysel for the many useful discussions which have contributed to the emergence of the view presented here. The main debt is to Gordon Tucker; together we have hammered out the sense in which we now interpret Halliday's original challenging concept of 'lexis as most delicate grammar'\u2022 I am also grateful to the three anonymous reviewers of this paper; the remaining infelicities being, as always, the author's. 2 The term 'typic' has been used in preference to 'generic' i n some of our earlier writings because we could give to this new term a has a clear definition -leaving 'generic' for 'generic (= genre) structure'. Here 'genetic' will serve equally well. 3 The fact that Leech could not propose a workable way to map the structures of language onto those that he inherited from standard logic is significant, but he can hardly be censured for failing to solve a problem that has also defeated all others who have attempted it. 4 There are occasions in natural language processing when it may be desirable and/or necessary to use system networks by working from right to left, e.g. as proposed (1) by Patten (1988) for some aspects of generation and (2) by O'Donoghue (1994) for the higher stages of parsing.",
	"cite_spans": [
	{
	"start": 1407,
	"end": 1420,
	"text": "Patten (1988)",
	"ref_id": "BIBREF24"
	}
	],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Notes",
	"sec_num": null
	},
	{
	"text": "But these are adaptations of the central concept for specific purposes, and they do not affect the main point that system networks are designed to be used in generation, working from left to right. 5 While Dahlgren has only 57 'category cuts', some of her terminal nodes such as 'animal' bring together a number of these and then lead off into further subsystems -and so effectively to further 'terminal nodes', such as 'vertebrate ' vs, non-vertebrate', and then 'mammal' vs. 'bird' vs.'fish' . Thus there are in practice a few more than 57 'terminal nodes'. But this does not affect the more general point being made here. 6",
	"cite_spans": [
	{
	"start": 432,
	"end": 493,
	"text": "' vs, non-vertebrate', and then 'mammal' vs. 'bird' vs.'fish'",
	"ref_id": null
	}
	],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Notes",
	"sec_num": null
	},
	{
	"text": "Our experience in COMMUNAL-working at both the semantic and the belief system levels -suggests that the development of even quite delicate areas of the network raises problems requiring a high level of expertise and experience for a satisfactory solution, and we believe that this experience is mirrored elsewhere. One can foresee that clients might be wise to subcontract the work of ontology-building back to the Penman team.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Notes",
	"sec_num": null
	},
	{
	"text": "Contrary to what is still a quite widespread assumption, there is no.need for a separate network from that for the nominal group to deal with the rank of the word; the nouns at the head of a nominal group are simply the realization of one part of its meaning (see Fawcett, Tucker and Lin 1993) . 8 These have ontological equivalents in the COMMUNAL belief system, so that interested persons may wish to compare this slicing of the universe of objects with those assumed for other ontologies. It would be an interesting exercise to get the creators of alternative taxonomies to explain their reasons for foregrounding their primary distinctions. 9 A pseudo-remedy that some systemicists use is to add complex wiring that prevents the realization of all but those combinations of features for which there is a realization (e.g. 'dog' plus 'female' but not 'hare' plus 'female'). But this to try to correct poor systemic modelling at too late a stage; it is the task of a system network to constrain possible co-selections. l0 The random generation of sentences is frowned upon by some NLG researchers, but it serves a valuable role in developing large generators, because it tests the availability of the lexicogrammatical resources, and so has a role in solving the 'expressibility' problem (Meteer 1992) .",
	"cite_spans": [
	{
	"start": 264,
	"end": 293,
	"text": "Fawcett, Tucker and Lin 1993)",
	"ref_id": "BIBREF10"
	},
	{
	"start": 1290,
	"end": 1303,
	"text": "(Meteer 1992)",
	"ref_id": "BIBREF20"
	}
	],
	"ref_spans": [],
	"eq_spans": [],
	"section": "7",
	"sec_num": null
	}
	],
	"back_matter": [],
	"bib_entries": {
	"BIBREF0": {
	"ref_id": "b0",
	"title": "Upper modelling: organizing language for natural language processing",
	"authors": [
	{
	"first": "J",
	"middle": [
	"A"
	],
	"last": "Bateman",
	"suffix": ""
	}
	],
	"year": 1990,
	"venue": "Procs of 5th International Workshop on Natural Language Generation",
	"volume": "",
	"issue": "",
	"pages": "54--61",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Bateman, J.A., 1990. 'Upper modelling: organizing language for natural language processing'. In Procs of 5th International Workshop on Natural Language Generation, Pittsburgh, pp. 54-61.",
	"links": null
	},
	"BIBREF1": {
	"ref_id": "b1",
	"title": "A General Organization of Knowledge for Natural Language Processing:the Penman Upper Model",
	"authors": [
	{
	"first": "J",
	"middle": [
	"A"
	],
	"last": "Bateman",
	"suffix": ""
	},
	{
	"first": "R",
	"middle": [
	"T"
	],
	"last": "Kaspar",
	"suffix": ""
	},
	{
	"first": "J",
	"middle": [
	"D"
	],
	"last": "Moore",
	"suffix": ""
	},
	{
	"first": "R",
	"middle": [
	"A"
	],
	"last": "Whitney",
	"suffix": ""
	}
	],
	"year": 1990,
	"venue": "USC/ISI Information Sciences Institute",
	"volume": "",
	"issue": "",
	"pages": "",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Bateman, J.A., Kaspar, R.T., Moore, J.D., and Whitney, R.A.. 1990. A General Organization of Knowledge for Natural Language Processing:the Penman Upper Model. Technical Report, USC/ISI Information Sciences Institute, Marina del Rey, California.",
	"links": null
	},
	"BIBREF2": {
	"ref_id": "b2",
	"title": "Introduction to Systemic Linguistics",
	"authors": [
	{
	"first": "M",
	"middle": [],
	"last": "Berry",
	"suffix": ""
	}
	],
	"year": 1977,
	"venue": "",
	"volume": "2",
	"issue": "",
	"pages": "",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Berry, M., 1977. Introduction to Systemic Linguistics, Vol 2: Levels and Links. London: Batsford.",
	"links": null
	},
	"BIBREF3": {
	"ref_id": "b3",
	"title": "Systemic Linguistics: Theory and Application",
	"authors": [
	{
	"first": "C",
	"middle": [
	"S"
	],
	"last": "Butler",
	"suffix": ""
	}
	],
	"year": 1985,
	"venue": "",
	"volume": "",
	"issue": "",
	"pages": "",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Butler, C.S., 1985. Systemic Linguistics: Theory and Applic- ation. London: Batsford.",
	"links": null
	},
	"BIBREF4": {
	"ref_id": "b4",
	"title": "Naive Semantics for Natural Language",
	"authors": [
	{
	"first": "K",
	"middle": [],
	"last": "Dahlgren",
	"suffix": ""
	}
	],
	"year": 1988,
	"venue": "",
	"volume": "",
	"issue": "",
	"pages": "",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Dahlgren, K, 1988. Naive Semantics for Natural Language. Boston: Kluwer.",
	"links": null
	},
	"BIBREF5": {
	"ref_id": "b5",
	"title": "Generating a sentence in systemic functional Grammar'. University College London (mimeo)",
	"authors": [
	{
	"first": "R",
	"middle": [
	"P"
	],
	"last": "Fawcett",
	"suffix": ""
	}
	],
	"year": 1973,
	"venue": "Readings in Systemic Linguistics",
	"volume": "",
	"issue": "",
	"pages": "146--83",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Fawcett, R.P., 1973/81. 'Generating a sentence in systemic functional Grammar'. University College London (mimeo). Reprinted in Halliday, M.A.K., and Martin, J.R., (eds.) 1981 , Readings in Systemic Linguistics. Batsford, pp. 146-83.",
	"links": null
	},
	"BIBREF6": {
	"ref_id": "b6",
	"title": "Cognitive Linguistics and Social Interaction: Towards an Integrated Model of a Systemic Functional Grammar and the Other Components of an Interacting Mind",
	"authors": [
	{
	"first": "R",
	"middle": [
	"P"
	],
	"last": "Fawcett",
	"suffix": ""
	}
	],
	"year": 1980,
	"venue": "",
	"volume": "",
	"issue": "",
	"pages": "",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Fawcett, R.P., 1980. Cognitive Linguistics and Social Interaction: Towards an Integrated Model of a Systemic Functional Grammar and the Other Components of an Interacting Mind. Heidelberg: Julius Groos.",
	"links": null
	},
	"BIBREF7": {
	"ref_id": "b7",
	"title": "The computer generation of speech with semantically and discoursally motivated intonation",
	"authors": [
	{
	"first": "R",
	"middle": [
	"P"
	],
	"last": "Fawcett",
	"suffix": ""
	}
	],
	"year": 1990,
	"venue": "Procs of 5th International Workshop on Natural Language Generation",
	"volume": "",
	"issue": "",
	"pages": "164--73",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Fawcett, R.P., 1990. 'The computer generation of speech with semantically and discoursally motivated intonation'. In Procs of 5th International Workshop on Natural Language Generation, Pittsburgh, pp. 164-73a.",
	"links": null
	},
	"BIBREF8": {
	"ref_id": "b8",
	"title": "Language as program: a reassessment of the nature of descriptive linguistics",
	"authors": [
	{
	"first": "R",
	"middle": [
	"P"
	],
	"last": "Fawcett",
	"suffix": ""
	}
	],
	"year": 1993,
	"venue": "Language Sciences",
	"volume": "14",
	"issue": "",
	"pages": "623--57",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Fawcett, R.P., 1993. 'Language as program: a reassessment of the nature of descriptive linguistics'. In Language Sciences 14.4: pp. 623-57.",
	"links": null
	},
	"BIBREF9": {
	"ref_id": "b9",
	"title": "A generationist approach to grammar reversibility in natural language processing",
	"authors": [
	{
	"first": "R",
	"middle": [
	"P"
	],
	"last": "Fawcett",
	"suffix": ""
	}
	],
	"year": 1994,
	"venue": "",
	"volume": "",
	"issue": "",
	"pages": "365--413",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Fawcett, R.P., 1994. 'A generationist approach to grammar reversibility in natural language processing'. In Strzalkowski, T., (ed.), Reversible Grammar in Natural Language Generation., Dordrecht: Kluwer, pp. 365-413.",
	"links": null
	},
	"BIBREF10": {
	"ref_id": "b10",
	"title": "How a systemic functional grammar works: the role of realization in realization",
	"authors": [
	{
	"first": "R",
	"middle": [
	"P"
	],
	"last": "Fawcett",
	"suffix": ""
	},
	{
	"first": "G",
	"middle": [
	"H"
	],
	"last": "Tucker",
	"suffix": ""
	},
	{
	"first": "Y",
	"middle": [
	"Q"
	],
	"last": "Lin",
	"suffix": ""
	}
	],
	"year": 1993,
	"venue": "New Concepts in Natural Language Generation",
	"volume": "",
	"issue": "",
	"pages": "114--86",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Fawcett, R.P., Tucker, G.H., .and Lin, Y.Q., 1993. 'How a systemic functional grammar works: the role of realization in realization'. In Horacek, H., and Zock, M., (eds.), 1993, New Concepts in Natural Language Generation, London: Pinter, pp. 114-86.",
	"links": null
	},
	"BIBREF11": {
	"ref_id": "b11",
	"title": "Categories of the theory of grammar",
	"authors": [
	{
	"first": "M",
	"middle": [
	"A K"
	],
	"last": "Halliday",
	"suffix": ""
	}
	],
	"year": 1961,
	"venue": "",
	"volume": "17",
	"issue": "",
	"pages": "241--92",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Halliday, M.A.K., 1961. 'Categories of the theory of grammar'. In Word 17, pp. 241-92.",
	"links": null
	},
	"BIBREF12": {
	"ref_id": "b12",
	"title": "The grammarian's dream: lexis as most delicate grammar",
	"authors": [
	{
	"first": "R",
	"middle": [],
	"last": "Hasan",
	"suffix": ""
	}
	],
	"year": 1987,
	"venue": "New Developments in Systemic Linguistics",
	"volume": "1",
	"issue": "",
	"pages": "184--211",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Hasan, R., 1987. 'The grammarian's dream: lexis as most delicate grammar'. In Halliday, M.A.K., and Fawcett, R.P. (eds.) 1987, New Developments in Systemic Linguistics, Vol 1: Theory and Description, London: Frances Pinter, pp. 184-211.",
	"links": null
	},
	"BIBREF13": {
	"ref_id": "b13",
	"title": "Motivating shared knowledge resources: an example from the PANGLOSS collaboration",
	"authors": [
	{
	"first": "E",
	"middle": [
	"H"
	],
	"last": "Hovy",
	"suffix": ""
	},
	{
	"first": "K",
	"middle": [],
	"last": "Knight",
	"suffix": ""
	}
	],
	"year": 1993,
	"venue": "Proceedings of the IJCAI Workshop on Shared Knowledge",
	"volume": "",
	"issue": "",
	"pages": "",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Hovy, E.H., and Knight, K., 1993. 'Motivating shared knowledge resources: an example from the PANGLOSS collaboration'. In Proceedings of the IJCAI Workshop on Shared Knowledge, Chambery, France.",
	"links": null
	},
	"BIBREF14": {
	"ref_id": "b14",
	"title": "Approximating an interlingua in a principled way",
	"authors": [
	{
	"first": "E",
	"middle": [
	"H"
	],
	"last": "Hovy",
	"suffix": ""
	},
	{
	"first": "S",
	"middle": [],
	"last": "Nirenburg",
	"suffix": ""
	}
	],
	"year": 1992,
	"venue": "Proceedings of the DARPA Speech and Natural Language Workshop",
	"volume": "",
	"issue": "",
	"pages": "",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Hovy, E.H., and Nirenburg, S., 1992. 'Approximating an interlingua in a principled way'. In Proceedings of the DARPA Speech and Natural Language Workshop, Hawthorne, NY.",
	"links": null
	},
	"BIBREF15": {
	"ref_id": "b15",
	"title": "Building a large ontology for machine translation",
	"authors": [
	{
	"first": "K",
	"middle": [],
	"last": "Knight",
	"suffix": ""
	}
	],
	"year": 1993,
	"venue": "Proceedings of DARPA Human Language Conference",
	"volume": "",
	"issue": "",
	"pages": "",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Knight, K., 1993. 'Building a large ontology for machine translation'. In Proceedings of DARPA Human Language Conference, March 1993.",
	"links": null
	},
	"BIBREF16": {
	"ref_id": "b16",
	"title": "Towards a Semantic Description of English",
	"authors": [
	{
	"first": "G",
	"middle": [
	"N"
	],
	"last": "Leech",
	"suffix": ""
	}
	],
	"year": 1969,
	"venue": "",
	"volume": "",
	"issue": "",
	"pages": "",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Leech, G.N., 1969. Towards a Semantic Description of English. London: Longman.",
	"links": null
	},
	"BIBREF17": {
	"ref_id": "b17",
	"title": "Longman Dictionary of Contemporary English",
	"authors": [
	{
	"first": "G",
	"middle": [
	"N"
	],
	"last": "Leech",
	"suffix": ""
	}
	],
	"year": 1974,
	"venue": "/81. Semantics. Harmondsworth: Penguin. Longman Group",
	"volume": "",
	"issue": "",
	"pages": "",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Leech, G.N., 1974/81. Semantics. Harmondsworth: Penguin. Longman Group, 1978. Longman Dictionary of Contemporary English. London: Longman.",
	"links": null
	},
	"BIBREF18": {
	"ref_id": "b18",
	"title": "Semantics (Vols 1 & 2). Cambridge: Cambridge",
	"authors": [
	{
	"first": "J",
	"middle": [],
	"last": "Lyons",
	"suffix": ""
	}
	],
	"year": 1977,
	"venue": "",
	"volume": "",
	"issue": "",
	"pages": "",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Lyons, J., 1977. Semantics (Vols 1 & 2). Cambridge: Cam- bridge University Press.",
	"links": null
	},
	"BIBREF19": {
	"ref_id": "b19",
	"title": "Text Generation and Systemic Functional Linguistics",
	"authors": [
	{
	"first": "C",
	"middle": [
	"M I M"
	],
	"last": "Matthiessen",
	"suffix": ""
	},
	{
	"first": "J",
	"middle": [
	"A"
	],
	"last": "Bateman",
	"suffix": ""
	}
	],
	"year": 1991,
	"venue": "",
	"volume": "",
	"issue": "",
	"pages": "",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Matthiessen, C.M.I.M., and Bateman, J.A., 1991. Text Genera- tion and Systemic Functional Linguistics. London: Pinter.",
	"links": null
	},
	"BIBREF20": {
	"ref_id": "b20",
	"title": "Expressibility and the Problem of Efficient Text Planning",
	"authors": [
	{
	"first": "M",
	"middle": [],
	"last": "Meteer",
	"suffix": ""
	}
	],
	"year": 1992,
	"venue": "",
	"volume": "",
	"issue": "",
	"pages": "",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Meteer, M., 1992. Expressibility and the Problem of Efficient Text Planning. London: Pinter.",
	"links": null
	},
	"BIBREF21": {
	"ref_id": "b21",
	"title": "Wordnet: an on-line lexical database",
	"authors": [
	{
	"first": "G",
	"middle": [],
	"last": "Miller",
	"suffix": ""
	}
	],
	"year": 1990,
	"venue": "International Journal of Lexicography",
	"volume": "3",
	"issue": "4",
	"pages": "",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Miller, G., 1990. 'Wordnet: an on-line lexical database'. In International Journal of Lexicography 3(4) (Special Issue).",
	"links": null
	},
	"BIBREF22": {
	"ref_id": "b22",
	"title": "Application-oriented computational semantics",
	"authors": [
	{
	"first": "S",
	"middle": [],
	"last": "Nirenburg",
	"suffix": ""
	},
	{
	"first": "C",
	"middle": [],
	"last": "Defrise",
	"suffix": ""
	}
	],
	"year": 1992,
	"venue": "Computational Linguistics and Formal Semantics",
	"volume": "",
	"issue": "",
	"pages": "",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Nirenburg, S., and Defrise, C., 1992. 'Application-oriented computational semantics'. In Johnson, R., and Johnson, M., (eds.) Computational Linguistics and Formal Semantics. Cambridge: Cambridge University Press.",
	"links": null
	},
	"BIBREF23": {
	"ref_id": "b23",
	"title": "Reversible Grammar in Natural Language Generation",
	"authors": [
	{
	"first": "T",
	"middle": [
	"F"
	],
	"last": "O'donoghue",
	"suffix": ""
	}
	],
	"year": 1994,
	"venue": "",
	"volume": "",
	"issue": "",
	"pages": "415--447",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "O'Donoghue, T.F., 1994. 'Semantic interpretation in a systemic grammar'. In Strzalkowski, T., (ed.), Reversible Grammar in Natural Language Generation., Dordrecht: Kluwer, pp. 415-447.",
	"links": null
	},
	"BIBREF24": {
	"ref_id": "b24",
	"title": "Systemic Text Generation as Problem Solving",
	"authors": [
	{
	"first": "T",
	"middle": [],
	"last": "Patten",
	"suffix": ""
	}
	],
	"year": 1988,
	"venue": "",
	"volume": "",
	"issue": "",
	"pages": "",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Patten, T., 1988. Systemic Text Generation as Problem Solving. Cambridge: Cambridge University Press.",
	"links": null
	},
	"BIBREF25": {
	"ref_id": "b25",
	"title": "A fast algorithm for the generation of referring expressions",
	"authors": [
	{
	"first": "E",
	"middle": [],
	"last": "Reiter",
	"suffix": ""
	},
	{
	"first": "R",
	"middle": [],
	"last": "Dale",
	"suffix": ""
	}
	],
	"year": 1992,
	"venue": "Proceedings of COLING-92: 14th International Conference on Computational Linguistics (Nantes). 1992. Morristown NJ: Bell Communications Research",
	"volume": "",
	"issue": "",
	"pages": "",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Reiter, E., and Dale, R., 1992. 'A fast algorithm for the generation of referring expressions'. In Proceedings of COLING-92: 14th International Conference on Computational Linguistics (Nantes). 1992. Morristown NJ: Bell Communications Research.",
	"links": null
	},
	"BIBREF26": {
	"ref_id": "b26",
	"title": "(eds.). Cognition and Categorization",
	"authors": [
	{
	"first": "E",
	"middle": [],
	"last": "Rosch",
	"suffix": ""
	}
	],
	"year": 1978,
	"venue": "",
	"volume": "",
	"issue": "",
	"pages": "",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Rosch, E., 1978. 'Principles of classification'. In Rosch, E., and Lloyd, B.B., (eds.). Cognition and Categorization. New Jersey: Erlbaum.",
	"links": null
	},
	"BIBREF27": {
	"ref_id": "b27",
	"title": "Understanding Natural Language",
	"authors": [
	{
	"first": "T",
	"middle": [],
	"last": "Winograd",
	"suffix": ""
	}
	],
	"year": 1972,
	"venue": "",
	"volume": "",
	"issue": "",
	"pages": "",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Winograd, T., 1972. Understanding Natural Language. Edinburgh: Edinburgh University Press.",
	"links": null
	}
	},
	"ref_entries": {
	"FIGREF0": {
	"uris": null,
	"type_str": "figure",
	"text": "Figure 2: A disjunctive entry condition",
	"num": null
	},
	"FIGREF1": {
	"uris": null,
	"type_str": "figure",
	"text": "Simultaneity in ontological relationships",
	"num": null
	}
	}
	}
	}