Automatic Question Classification For Speech Based Amharic Question Answering

Abstract

A question answering system is one of the disciplines under an information retrieval that displays precise expected answers from the huge documents for a specific question. The existing QA system doesn’t support visually impaired Amharic monolingual speakers. Hence, those users cannot able to access the required information. Therefore, this study focused on developing an interactive interface using both Amharic speech recognition and synthesizer for factoid question answering with an automatic question classification. Besides, it gives an emphasis particularly on designing audio application for visually impaired Amharic monolingual speakers. This study is conducted for designing and constructing to automatic question classification for speech-based Amharic question answering. To this end, concatenate phoneme unit selection methodology is used for speech synthesis and SVM for question classification. After all, speech recognition, question answering, and speech synthesizer are combined to construct the study. Accordingly, various tools were using in order to construct a prototype of the system. For speech recognition, sphinx4 decoding tool; for question answering Lucene, LibSVM, and for developing the whole system, java NetBean 7.1.1 were used. We have used 22600 news articles from different newspapers (Ethiopian News Agency, Ethiopian Reporter, and from the cloud) which are prepared in 4542 files for training and testing. In this study, used 2,016 speech question sentences corpus by 24 (9 Female and 15 Male) different peoples, each having read 84 question spoken words and the questions are numeric and person related. The experimental results of speech recognition system achieved 85.58% of accuracy. Furthermore, the speech synthesis correctly pronounced 80.86% with 3.17 and 3.45 accuracy in; intelligibility and naturalness based on MOS. In addition, the SVM question classification provides 73.91% precision and 94.44% recall and 82.92%F-measure. In general, the speech-based Amharic question answering system achieves 72.75%. The challenge of these study is, it didn't use a synonyms words to parsing a query. Therefore, as recommendation designing and developing semantic similarity using ontology based structure is needed to enhance the performance of Speech based Amharic question answering system.