Classifying Questions in Question Answering System Using Finite State Machines with a Simple Learning Approach

Abstract

Question Classification plays a significant part in Question Answering system. In order to obtain a classifier, we present in this paper a pragmatic approach that utilizes simple sentence structures observed and learned from the question sentence patterns, trains a set of Finite State Machines (FSM) based on keywords appearing in the sentences and uses the trained FSMs to classify various questions to their relevant classes. Although, questions can be placed using various syntactic structures and keywords, we have carefully observed that this variation is within a small finite limit and can be traced down using a limited number of FSMs and a simple semantic understanding instead of using complex semantic analysis. WordNet semantic meaning of various keywords to extend the FSMs capability to accept a wide variety of wording used in the questions. Various kinds of questions written in English language and belonging to diverse classes from the Conference and Labs of the Evaluation Forum’s Question Answering track are used for the training purpose and a separate set of questions from the same track is used for analyzing the FSMs competence to map the questions to one of the recognizable classes. With the use of learning strategies and application of simple voting functions along with training the weights for the keywords appearing in the questions, we have managed to achieve a classification accuracy as high as 94%. The system was trained by placing questions in various orders to see if the system built up from those orders have any subtle impact on the accuracy rate. The usability of this approach lies in its simplicity and yet it performs well to cope up with various sentence patterns.