Atkins, B. T., Kegl, J. and Levin, B. 1988 Anatomy of a Verb Entry. International Journal of Lexicography 1(2):84—126.

Boguraev, B. K., and Briscoe, E. J. 1989 Computational Lexicography for Natural Language Processing. Longman, London and New York.

Church, K. W. and Hanks, P. 1989 Word Association Norms, Mutual Information, and Lexicography. Proceedings of the 27th Annual Meet­ing of the Associaton for Computational Linguistics, Vancouver, B.C.

Fillmore, C. J. 1989 Two Dictionaries. International Journal of Lexicog­raphy 2(l):57-83.

Hindle, D. 1989 Acquiring Disambiguation Rules from Text. Proceedings of the 27th Annual Meeting of the Association for Computational Linguistics, Vancouver, B.C.

Landau, S. I. 1989 Dictionaries: The Art and Craft of Lexicography. Cambridge University Press, Cambridge, U.K.

Branimir Boguraev is a Research Staff Member at the IBM T.J. Watson Research Center, where he is also in charge of the Lexical Systems Project. He holds a Ph.D. degree from the University of Cambridge, where he subsequently worked on a number of projects in natural language processing. His recent work has been in the areas of computational lexicography and lexicology. Boguraev's address is: IBM Research, P.O. Box 704, Yorktown Heights, New York 10598. E-mail: bkb@ibm.com

(Information Sciences Institute, University of Southern California) Reviewed by Wolfgang Hoeppner University of Koblenz

This book is a revised version of the author's dissertation, which was submitted to Yale University in February 1987 and published as a research report (Hovy 1987). One shouldn't be too surprised to learn from the preface that Roger C. Schank, Drew McDermott, and Bob Abelson were the honorable members of the thesis committee. Vari­ous well-known AI researchers—not exclusively from Yale—have given a hand while the thesis was on its way and are therefore mentioned in the acknowledgments. The acknowledgments, by the way, give a first example of how stylistic features affect the generation of text: Hovy switches between several styles (formality, verbosity, gratefulness, haste) while expressing his thanks to different classes of persons. The preface stresses that the book is not only useful to computational linguists, but also to theoretical linguists, especially those working in generation. Thus, the last section of every chapter deals with implementation and might be skipped by readers not interested in the computa­tional issues.

The book describes the generation system PAULINE, which was developed and implemented by the author. The sys­tem's name is an acronym: Planning And Uttering Lan­guage In Natural Environments. (It is also the name of Hovy's sister).

Chapter 1(11 pp.) introduces the specific research area: how do pragmatic and stylistic issues influence the genera­tion of natural language texts? Starting with real-world descriptions of an event at Yale, the destruction of a shantytown by university authorities, it is demonstrated how different viewpoints of the respective authors affect texts. The same event is then described by Pauline in different pragmatic adjustments. The event itself is repre­sented in a network of about 120 elements (presumably some version of conceptual dependency), and it is claimed that the system produces over 100 different texts. The second example is a description of a fictitious primary election between Carter and Kennedy as Democratic presi­dential candidates. "Well, so Carter lost the primary to Kennedy by 1335 votes" is one example of a very condensed description. This terseness is beaten only by example num­ber 12, which consists of nothing but blanks: "The program didn't find any topics that it liked and the hearer also liked,

and it didn't have time to search the story representation for other topics or to find ways of mitigating the unpleasant ones" (p. 8). The third event deals with a fight between Mike and Jim, resulting in the death of the latter. All three events are used throughout the book. I didn't count the different versions of the texts generated by the system and included as examples in the book, but one gets a little bored when reading, e.g., the nth version of the shantytown story.

The programmatic essence of Hovy's approach can best be characterized by the following quotation:

Current artificial intelligence work in natural language process­ing places far too little emphasis on the role of pragmatics. This is a mistake. ... The time is ripe to start examining what kinds of goals are relevant, what strategies achieve them, and what planners and realizers must be like to operate under their control. We must face the fact that, in order to have real, flexible text, we simply cannot do without recourse to the airy world of pragmatics. We need to experiment! (p. 11)

True enough!

Chapter 2 (23 pp.) is briefly called "Pragmatics," and contains the theoretical underpinnings of Pauline. Start­ing with a short discussion on what pragmatics is all about and what relation to semantics might be established, Hovy identifies three basic categories that are relevant to his system (p. 17):

• Interlocutor's personal characteristics: factual knowl­edge, opinions, emotional states, interpersonal relation­ship, etc.

• The speaker's goals with respect to the hearer: effects on future behavior, opinions, relative status, etc.

• The conversational atmosphere: tone, time, physical set­ting, etc.

Each of these categories is illustrated with several examples in subsequent sections. This leads to a set of 13 goals, which are a first approximation to the underlying pragmatic features necessary for generation. Some examples of these goals are: increase knowledge; make the topic seem good or bad, contrary to the hearer's opinion; make the hearer feel inferior to, equal with, or dominant over the speaker; be hasty, normal, or effusive. Hovy does not claim that these goals are complete or adequate in a general way. They rather appear to be necessary for the way pauline is supposed to produce text. One might wonder how a system would be able to choose rationally among these goals, and this, of course, is beyond the scope of an implemented system. The user has to determine the interpersonal goals and the conversational setting by selecting a value for 23 relevant features, each of these being represented by three possible values. For instance, the emotional state of the hearer can be set to happy, angry, or calm; the conversa­tional atmosphere with respect to time can be set to much time, some time, or little time. Rhetorical goals relate the pragmatic adjustments defined by the user, or rather the experimenter, to decisions for the generator, e.g., word choice, sentence inclusion. The chapter closes with a discussion of rhetorical goals of opinion and of style. The latter has seldom been investigated in computational linguistics (exceptions are the EPISTLE project [Heidorn et al. 1982], or the BOGUE system [Ryan 1989]), and it is certainly a merit of Hovy's work to establish an initial set of 12 style features (e.g., timidity, haste, respect).

In Chapter 3(17 pp.), Hovy argues for an interpretation component as part of a natural language generator. Inter­pretation in this sense means that a set of inference rules has to be present to "digest" the conceptual input represen­tation and produce a structure that is suited for the task of verbalization. Examples for such inference rules are the determination of an appropriate level of detail, presenta­tion in a confrontative or conciliatory manner, or the inclusion of remindings. In pauline, the interpretation mechanism is realized as patterns that change the original network, when matched with some input configuration.

Chapter 4 (25 pp.) discusses the notion of affect in texts, i.e., the opinion a speaker has about the content of his or her utterances. The effect of such a biased generation strategy is demonstrated with different issues of the fighting scene between Jim and Mike, where the differences originate from the agent the system sides with. Hovy develops an "affect rule": "In order to convince the hearer that some topic is GOOD or BAD, combine it with other GOOD or BAD topics using enhancers and mitigators" (p. 61). This rule appeals to both aspects of the generation process. Content is determined by the selection of topics and their linearization. Three plans are provided that have a similar function to the text schemata used in the TEXT system (McKeown 1985). These plans are called DESCRIBE, RELATE, and CONVINCE, the last one being presented in some detail, as this is the one where affect sneaks in. The production of form is influenced by affect-bearing phrases (e.g. "not only X but Y"), by adverbs (e.g. "really," "only," "slightly"), by clause order, and by word choice (e.g. "wimpy" versus "small"). Chapter 4 closes with some observations concerning partiality of the generated text. The rhetorical goals "timidity" and "partiality," or rather their respective values, are calculated on the basis of other features.

Chapter 5 (23 pp.) is about style. Hovy introduces three style notions: formality, haste, and force. For each of these rhetorical goals, some rules are given that allow the system to derive values by combining already established rhetori­cal goals. "RG:haste", for example, "is set to 'pressured' if the time is marked 'little', the relative social status is marked 'subordinate', and the depth of acquaintance is marked 'acquaintances' or 'strangers' " (p. 93). Such an "algorithmic approach to the creation of style" is what leads to the diversity of texts generated by pauline and is motivated by the assumption "that style is the result of following a coherent policy" (p. 104).

In Chapter 6 (24 pp.), Hovy argues for the central role of the lexicon as the one and only device for linguistic knowl­edge within a generator. This, obviously, is quite contrary to current grammatical theories, as for instance the varie-

ties of unification-based approaches. Hovy briefly sketches issues in transformational, systemic, and functional gram­mar, and stresses the need for a "phrasal lexicon" as originally introduced by Becker (1975). He then introduces the "syntax specialists," which incorporate the three tasks of inclusion of topics, ordering, and casting (selection of syntactic classes) in a procedural manner. This approach resembles the one earlier developed for analysis and known as "word-expert parsing" (Small and Rieger 1982). This analogy, however, is an inference of the reviewer and is not discussed in the book. The last section of Chapter 6 illus­trates the incremental generation of the sentence "The small 23-year-old from New Haven, Sue, was told by Jim that Janet died" as a realization of a conceptual depen­dency representation of something like "Jim told Sue that Janet died." Appendix B contains the complete phrasal grammar implemented for pauline.

Chapter 7 (17 pp.) once more turns to the planning process incorporated in pauline. Hovy distinguishes be­tween prescriptive planning and restrictive planning, the first of these methods being the one used in generators up to now. He argues for the incorporation of both techniques, which is called "limited-commitment or interleaved planning." The essence of this method of planning is that there is no fixed and predefined order between the planner (expansion of states) and the realizer (testing whether a planned action has contributed to the ultimate goal). Thus results of the realizer should be able to improve the planner at all times during the planning process. Hovy discusses planning approaches in other generators, e.g., KAMP (Appelt 1985), the system which is best known for its extreme plan-based philosophy. The last section of this chapter describes the interleaved planning employed by pauline.

The book's last chapter contains a nine-page section, "Review of Language Generation" a title that is not fully justified as it is with McKeown and Swartout (1988) or McDonald (1987). The main contribution of this chapter is the addition of one fundamental question for generation to the already existing two questions. Hovy adds to 'What should I say?' and 'How should I say it?' the question 'Why should I say it?', thus manifesting his pragmatic approach. A question, but quite a different one, has always puzzled me with American dissertations: why is the state of the art always contained in the last chapter, unlike European dissertations, where this is usually the starting point? For my part, I have developed the habit of starting to read an American dissertation at the last chapter and then continu­ing at the beginning.