Generating Adjectives to Express the Speaker's Argumentative Intent

being modified. In addition, these decisions interact with the lexical properties of adjectives, the syntax of the clause We address the problem of generating adjectives in a text and other factors like collocations. In this paper we theregeneration system. We distinguish between usages of adfore address the following two questions: What should be jectives informing the hearer of a property of an object and the input to a generator capable of producing argumenusages expressing an intention of the speaker, or an artative usages of adjectives? And how should the generator gumentative orientation. For such argumentative usages, combine the many interacting factors constraining the we claim that a generator cannot simply map from inforselection of an adjective? mation in the knowledge base to adjectives. Instead, we After reviewing previous work related to these quesidentify various knowledge sources necessary to decide tions, we present the linguistic data upon which we base whether to use an adjective, what adjective should be our approach and the conclusions we draw from its selected and what syntactic function it should have. We analysis. We then present and justify the input we require show how these decisions interact with lexical properties of to properly select adjectives and discuss how adjective adjectives and the syntax of the clause. We propose a selection is constrained by the lexical properties of adjecmechanism for adjective selection and illustrate it in the tives and interacts with other surface decisions. The paper context of the explanation component of the ADVISOR expert illustrates the key features of our implementation of adjecsystem. We describe an implementation of adjective selective selection in the context of the ADVISOR explanation tion using a version of Functional Unification Grammars. component. Introduction Previous Work Traditionally, an adjective is defined as ‘‘serving as a In previous work in generation, adjectives have been modifier of a noun to denote a quality of the thing named, studied as a tool for producing descriptions of objects. It is to indicate its quantity or extent, or to specify a thing as important to distinguish usages of descriptive noun-phrases distinct from something else’’ (Webster, 1963). Analysis to either refer to objects or to attribute a property to objects of human conversations however shows that adjectives of(Donnellan, 1966, Kronfeld, 1981, Searle, 1979). In a ten loosely relate to actual properties of the objects being referential usage, a noun-phrase is used when the speaker modified but are used to express a speaker’s intention or wants the hearer to identify some object. In this case, argumentative orientation. The work we present here is adjectives are used to contrast the target object from other developed in the context of the explanation component of potential referents. The proper adjectives are chosen based the ADVISOR expert system (McKeown et al, 1985, on their discriminatory power. For example, in a backMcKeown, 1988), a question answering system advising ground containing blocks of different forms and colors, the university students which courses to select. In this context, generator will pick a combination of form and color that when an academic advisor tells a student that a course is can be used to uniquely identify the referent and differenvery hard, he often does not refer to a property of the tiate it from all other blocks in the background. Different course, but rather expresses his evaluation of the course. mechanisms for such a selection are presented in (Dale, This creates problems for text generation. The first 1988, pp.249-262) for the EPICURE system, (Appelt, problem we face is that the information needed to choose 1985) for the KAMP system and in (Reiter, 1990). whether to use an adjective playing an argumentative role In attributive usages, the goal of the speaker is to inform cannot be found directly in a knowledge-base describing the hearer of some property of an object. In (McKeown, objects of the domain. Instead, the decisions must be 1985) and (Appelt, 1985) for example, adjectives are used based on the speaker’s goals, a hearer model and the object to perform inform speech-acts. In this case, the generator 1Reprinted from Proceedings of the 9th National Conference on Artificial Intelligence (AAAI 91), 1991, Anaheim CA, pp98-104 simply maps from the information in the knowledge-base (McKeown & Elhadad, 1991, Elhadad & McKeown, describing the object to an adjective denoting the property 1990). We identified a set of pragmatic features necessary being attributed. Note that in KAMP (Appelt, 1985), the to distinguish between these connectives, including arnotion of action subsumption was introduced to account for gumentative features. In this paper, we refine this work cases where a particular noun-phrase simultaneously and identify features to adequately select a certain class of served as a referring and attributive expression and the adjectives. adjective was selected both because of its contrastive value and of its informative value. The Problem: Data and Motivation Other works have studied usages that are neither attribuOriginally, our task was to extend the linguistic coverage tive nor referential in the sense discussed above. With of the generator for the explanation component of the PAULINE, Hovy (Hovy, 1988) discussed the use of adjecADVISOR expert system to select adjectives based on tives to satisfy pragmatic constraints. For example, the general principles. ADVISOR is a system that assists univergenerator could produce a sentence like poor John was sity students select courses and plan their semester severely beaten by the police where poor does not denote (McKeown, 1988). any information about John but rather expresses the orienWe performed an analysis of a corpus of 40,000 words tation of the speaker. Hovy covered many different lincontaining transcripts of recordings of advising sessions guistic devices satisfying pragmatic constraints and as a with human academic advisors. In this corpus, we idenresult provides only a very superficial treatment of adjectified approximately 700 occurrences of 150 distinct adjective selection (he devotes a single paragraph to its discustives. We focused our analysis on all occurrences of adjecsion). tives modifying a course, in both predicative and attribuIn (Bruxelles et al, 1989) and (Bruxelles & Raccah, tive positions. We found 69 such occurrences, of 26 dis1991), a model for describing the argumentative potential tinct adjectives. Figure 1 shows a break down of these of lexical items is introduced. This model aims at explainoccurrences in semantic classes. ing how adjectives like courageous express both a property of the modified object and an argumentative orientation of Of the 69 occurrences listed in Figure 1, 58 express a the speaker (a favorable evaluation of the object), whereas property of a course that one cannot reasonably expect to adjectives like intrepid or bold while conveying roughly find in the knowledge-base describing courses. For exthe same information also convey a different orientation. ample, it is problematic to describe a course as good or The reported work is still at early stages and is oriented hard in absolute terms. For most of the occurrences theretowards interpretation. We use here many concepts fore, the technique of mapping from a semantic property in derived from this work and examine its implications on the knowledge-base to an adjective, as used in previous generation. generation systems to produce attributive noun-phrases, In earlier work, we have studied the problem of generatwould not be applicable. Most of the usages of adjectives ing certain connectives like but, although, because or since in the corpus correspond to an argumentative usage. For Semantic class Adjective Occurrences Semantic class Adjective Occurrences Difficulty [24] advanced 1 Importance [24] important 10