Natural language generation has received much less scrutiny than other aspects of language processing. McKeown is one of a small set of recent researchers to focus their attention in this direction. The book is a description of her work, based largely on her PhD thesis research.
Despite the title, this book is neither an overview of, nor an introduction to, generation. Rather, it is a report on McKeown’s own area of specialization: discourse structure (i.e., a grammar of exposition) and coherency (not shifting focus arbitrarily). She presents a theory of rhetoric that might be derived from style manuals prescribing formulaic strategies for expository writing. She classifies sentences into ten types, and introduces four discourse structures and three discourse goals. These constructs are used in a system that answers simple questions about entities in a database model. McKeown proposes the use of focus not only to order sentences, but also to select the facts to state. The program, TEXT, outputs paragraphs consisting of (well-organized) lists of facts.
Her system operates as follows:
(1) It receives a query.
(2) It fetches the schema associated with the corresponding discourse structure.
(3) It selects the relevant subnet of the database model.
(4) It calls the procedures listed in the schema to retrieve some facts to say, using focus constraints to prevent jumpiness.
(5) It puts the out-filled schema through a dictionary to get a deep structure (following David McDonald’s approach [1]).
(6) It transforms the deep structure to English (using techniques developed by Martin Kay [2]).
McKeown also proposes a discourse history mechanism, to prevent repetition of facts. The interaction between the discourse structure and the database model is implemented with ATNs, a message formalism, matching, and function calls. McKeown suggests that other questions and other types of text can be handled by adding more discourse structures, and slightly extending her program. She suggests that the big open problem is an explicit reader model (the commonly accepted cure-all). In passing, she advocates development of a real knowledge representation with real inference.
While McKeown has selected an important topic for treatment, in the end the result is unsatisfying. Her work is often vague or unclear precisely where it is most intriguing. She tends to present the material without sufficient analysis or motivation. For example, her taxonomies are defined too vaguely, and are not well motivated. The description of the interaction of the discourse structure and the database model is likewise unclear. For example, she observes that backtracking was not a problem in practice, but doesn’t suggest why this is so. The general discussion of text generation (Chapters 1 and 6) describes the conventional wisdom on the subject (e.g., the strategy/tactics dichotomy), but does not explain or justify it. While McKeown has integrated ideas from McDonald [1], Kay [2], Sidner [3], etc. into a complete system, she never explains why these methods were good choices. She claims that the bottleneck of the system was the “tedious” encoding of the maps from knowledge nodes to English words and syntax. Yet she brushes off this problem as devoid of “theoretical value.”
The book itself suffers from a poor typeface. There are also a few relics of incomplete conversion from earlier writings.
Perhaps McKeown’s most important contribution is the integration of components to form a complete system. This is a nontrivial task. While McKeown’s explanations may leave something to be desired, the book is a useful place to look at some of the difficulties involved in the construction of a real natural language system.
