Syntax-driven argument identification and multi-argument classification for semantic role labeling

Abstract

Semantic role labeling is an important stage in systems for Natural Language Understanding. The basic problem is one of identifying who did what to whom for each predicate in a sentence. Thus labeling is a two-step process: identify constituent phrases that are arguments to a predicate, then label those arguments with appropriate thematic roles.

Existing systems for semantic role labeling use machine learning methods to assign roles one-at-a-time to candidate arguments. There are several drawbacks to this general approach. First, more than one candidate can be assigned the same role, which is undesirable. Second, the search for each candidate argument is exponential with respect to the number of words in the sentence. Third, single-role assignment cannot take advantage of dependencies known to exist between semantic roles of predicate arguments, such as their relative juxtaposition. And fourth, execution times for existing algorithm are excessive, making them unsuitable for real-time use.

This thesis seeks to obviate these problems by approaching semantic role labeling as a multi-argument classification process. It observes that the only valid arguments to a predicate are unembedded constituent phrases that do not overlap that predicate. Given that semantic role labeling occurs after parsing, this thesis proposes an algorithm that systematically traverses the parse tree when looking for arguments, thereby eliminating the vast majority of impossible candidates.

Moreover, instead of assigning semantic roles one at a time, an algorithm is proposed to assign all labels simultaneously; leveraging dependencies between roles and eliminating the problem of duplicate assignment.

Experimental results are provided as evidence to show that a combination of the proposed argument identification and multi-argument classification algorithms outperforms all existing systems that use the same syntactic information.