A linguistic approach to temporal information analysis
Richard Sheng · 1984
This dissertation presents linguistic approach, based on test-score semantics and fuzzy logic, to the problem of fuzzy temporal inference in question answering systems. While several models have been proposed for the analysis of historical information, little research has been done satisfactorily on the manipulation of fuzzy temporal quantifiers such as a few days ago, which refers to absolute count, and often, which refers to relative count. In this dissertation, the theory of possibility is taken as an appropriate framework for dealing with such problem. The first step of the approach for statements with absolute-count temporal quantifiers only, is to translate these statements into their corresponding possibility assignment equations using test-score semantics. Time information among events can then be inferred from these equations by deriving its possibility distribution. The compositional rule of inference plays the central role in solving the equations. The algorithm for computing the compositional rule can be transformed into one similar to Gaussian elimination for the solution of linear equations. This method is straightforward but inefficient. An improved technique is proposed which represents all quantifiers by fuzzy numbers and replaces the rule of composition with fuzzy addition. With this technique the desired possibility distribution can be obtained by applying generalized Warshall's algorithm on the Time-Relation-Matrix, which is representation for given information. As for quantifiers which refer to relative count, such as often, fuzzy multiplication will play the primary role. Finally, all mechanisms amount to fuzzy temporal logic, which serves as an inference tool for temporal information analysis. This computational approach to fuzzy temporal quantifiers has the following advantages. (1) It subsumes those approaches based on two valued logic as limiting cases. (2) It deals uniformly with the relative and absolute time relations. (3) It is easy to implement. (4) It can ge generalized to apply to other quantifiers used in different concepts such as temperature and distance.