A Hybrid Approach to Inferring a Consistent Temporal Relation Set in Natural Language Text

Abstract

This dissertation investigates the temporal relation identification

task. The goal is to construct consistent temporal relations between

temporal entities (e.g., events and time expressions) in a

narrative. Constructing consistent temporal relations is challenging

due to the exponential increase in the number of candidates for

temporal relations proportional to the number of pairs of temporal

entities. When we use transitive constraints to construct consistent

temporal relations, performance improvement can be expected because

the application of transitive constraints reduces the number of

possible relation candidates in a narrative.

The primary objective of this study is to develop a temporal relation

identification (TRI) system that is composed of three modules: 1) a

module that classifies the temporal relation of a pair of temporal

entities, 2) a module that extracts conflicting classified relations

using transitive constraints, and 3) a module that restores consistent

temporal relations from the conflicting relations using transitive

constraints. In developing a TRI system, this dissertation examines

whether the application of transitive constraints to such a system can

lead to performance improvement.

The first step in developing the system was to implement a rudimentary

temporal relation classification module. The module labels a pair of

temporal entities with a temporal relation among eleven possible

temporal relations. Next, a method for extracting conflicting

relations among classified relations is proposed. The extraction

method is based on heuristics because of NP-hard complexity in

extracting all conflicting relations. Finally, two heuristic methods

are proposed that restore consistent temporal structure from

conflicting relations using transitive constraints. The performance of

the developed system is tested using TimeBank and AQUAINT temporal

corpora.

The results of this work indicate that a performance improvement

through the application of transitive constraints to TRI task is not

guaranteed. Furthermore, this study empirically shows the limitations

on performance improvement through the application of transitive

constraints to the TRI task and identifies the bottlenecks in the TRI

task.