DFKI-LT - Dissertation SeriesVol. XII
Johannes Bos: Underspecification and Resolution in Discourse Semantics
price: € 13
Several natural language expressions are known to be ambiguous, that is, there are different ways in which we can interpret them. Although human beings seem easily able to process these ambiguities (it is mostly the context that helps us to get the right interpretation), it is enormously difficult to tell machines how to deal with this phenomenon. The aim of this thesis is to design and implement a formalism for processing such natural language ambiguities.
In Part I we develop a formalism for describing semantic ambiguities, or to use modern terminology: underspecification. The semantic representations used in this formalism describe natural language expressions as ambiguous semantic representations. Resolving these representations leads to a set of unambiguous semantic representations. We mainly deal with scope ambiguities, and show how to apply the method, known as Hole Semantics, to predicate logic.
An attractive property of Hole Semantics is that it is independent of the logical forms it describes. For instance, to deal with context-sensitive expressions such as pronouns and presuppositions, Discourse Representation Theory (DRT) has been proposed. DRT works with Discourse Representation Structures (DRSs), and we show how to combine Hole Semantics with DRSs. Moreover, we show that Hole Semantics can play a role in practical natural language processing applications, in particular machine translation.
Part II is devoted to resolving underspecified representations of Hole Semantics. We present and compare three different ways of resolving scope ambiguities. The first approach uses model building for first-order logic as a technique to resolve scope ambiguities. Because we use off-the-shelf general purpose reasoning tools in this approach, this is an attractive option. Nonetheless, the processing times are too meager for practical use (at least for the time being - this might change in the future, as model building is a strongly developing field). The second approach implements a resolution algorithm in the programming language Prolog, showing better performance. Both of these methods generate in principal all solutions after the construction of the underspecified representation. The third approach, however, produces one (default) solution during semantic construction, based on linguistic information.
Although most of the time and space is devoted to scope ambiguities, the formalism is extended to cover presuppositional expressions. We reformalize and implement Van der Sandt's projection algorithm for presuppositions. The nature of this algorithm is generate-and-test, where solutions are disregarded when they do not meet certain maxims of conversation, known as the acceptability constraints. Verifying whether these constraints are violated, requires a great deal of inference. By translating DRSs to first-order formulas, we show that state-of-the-art theorem provers for first-order logic can serve well to perform these inferences.