Computer Science Colloquium, 2007-2008

Title: Packed Computation of Exact Meaning Representations
 

Abstract:
An important question in Natural Language Understanding (NLU) is how to improve accuracy in NLU tasks. Accuracy is paramount is "exact NLU" applications, such as solving word problems (logic puzzles, math/physics/chemistry questions), understanding regulatory texts and controlled language, as well as NL interfaces to databases. These applications require exact meaning representations that rely on knowledge of structural semantics -- the meaning of functional words (quantifiers, connectives, comparatives, etc.) and how they affect the meaning of sentences. Exact meaning representations allow the computer to accurately capture and integrate the information that appears throughout the document and to draw appropriate inferences from it. Even in other NLU applications such as question answering, using knowledge of structural semantics could improve the accuracy of understanding functional words and of information integration.

Three main questions pertaining to exact meaning representations are: 1) How can the representations be calculated given one syntactic analysis of a sentence? 2) How can all possible representations be calculated efficiently given a packed syntactic analysis (parse forest)? 3) How can the coverage of semantic analysis be extended to additional linguistic constructions?

In my dissertation, I address these three questions. I show how the syntax-to-semantics mapping can be specified more easily than in traditional approaches by using the framework of Glue Semantics (linear logic). I then develop a novel algorithm that efficiently computes a packed meaning representation given a packed syntactic analysis -- this combines the framework of Glue Semantics with the general framework for ambiguity management developed at the Palo Alto Research Center ("choice-space packing"). In the second half of the dissertation, I extend the coverage of semantic analysis to advanced linguistic constructions, including comparatives, reciprocals, and words such as 'same' and 'different', where the mapping from syntax to semantics is complex. In my talk, I will mainly focus on the algorithm (and necessary background).