Realistic Language Comprehension
Christopher K. Riesbeck Yale University
Natural-language processing by computer, particularly the comprehension of natural-language texts, has been the subject of much work and even more debate in Artificial Intelligence (AI) circles. A long-running battle has been fought over whether an analyzer should look like Fig. 2.1 or Fig. 2.2.
Another set of arguments has been about the nature of the meaning representation produced by the analyzer. Should it be procedural or declarative, or doesn't it matter? Should it use primitives (e.g., Conceptual Dependency) or patterns of interrelationships (e.g., KRL) to define and distinguish concepts? At Yale, we believe in the simultaneous application of semantic and syntactic knowledge (i.e., the second flow chart), and we use primitives in a declarative framework of meaning representation.
But I want to avoid debate here about the best way to construct an analyzer, and the best way to represent meaning. Instead, I would like to look at what constitutes reasonable input material for building and testing languageunderstanding models. It is becoming increasingly obvious in the field that we have to deal with real texts, texts that were originally generated to communicate, not to test parsers. The days when we can compare programs by how well they handle "Max went to the store" have passed. In fact, the days of dealing with single sentences have pretty much disappeared.
This change has come about for several reasons. First, despite many false starts, there has been real progress in natural-language understanding. We are now writing programs to deal with multisentential connected texts, rather than