Rule Generation Based On Structural Clustering For Automatic Question Answering

Loading...
Thumbnail Image
Date
2009-12
Authors
Song, Shen
Journal Title
Journal ISSN
Volume Title
Publisher
Universiti Sains Malaysia
Abstract
In rule-based methods for Question-Answering (QA) research, typical rule discovery techniques are based on structural pattern overlapping and lexical information. These usually result in rules that may require further interpretation and rules that may be redundant. To address these issues, an automatic structural rule generation algorithm is presented via clustering, where a center sentence-based clustering method is designed to automatically generate rules for QA systems. The methodology for this research involves three phases. The first phase involves pre-processing of training question-answer pairs derived from the Canadian Broadcasting Corporation’s (CBC) 4 Kids reading comprehension corpus. Pre-processing also involves part-of-speech (POS) tagging. The second phase involves automatic rule generation where the POS-tagged QA pairs are clustered based on the similarity in matching POS tokens and their sequences. For this, the BLEU similarity computation method is employed. The final phase involves the operationalisation of the QA system called Question Answering System based on Automatic Rule Generation (QASARG). The output from this system is then evaluated. The effectiveness of QASARG was evaluated against another rule-based QA system, Quarc. The accuracy of QASARG is in the range of 55% to 85% depending on the question type, and these are on average 26.4 % higher than those for Quarc. However, it must be noted that the test data sets used to evaluate QASARG and Quarc are different (i.e. QASARG is tested based on question-answer pairs derived from the reading comprehension passage while Quarc’s results are based on the entire reading comprehension passages). Nevertheless, the results for QASARG indicate that structural similarities between sentences are useful in generating reliable and reasonably accurate rules for QA systems.
Description
Keywords
Rule generation based on structural clustering , for automatic Question answering
Citation