Study and Implementation of Monolingual Approach on Indonesian Question Answering for Factoid and Non-Factoid Question

Study and Implementation of Monolingual Approach on Indonesian Question Answering for Factoid and Non-Factoid Question ∗∗∗∗ Alvin Andhika Zulen a and Ayu Purwarianti b School of Electrical Engineering and Informatics, Bandung Institute of Technology, Ganesha 10, Bandung 40135, West Java, Indonesia a alvin.andhika@gmail.com b ayu@stei.itb.ac.id Abstract. We developed an open domain QA system that can handle factoid and non- factoid questions in Indonesian language by using monolingual approaches. EAT classification is done by identifying question word and clue words. Keyword extraction from question is done by looking at POS information of each word in question, eliminating stop words, and stemming. We use articles from Indonesian Wikipedia as corpus and Lucene framework as the base for passage retriever component, with three additional processing: query expansion, boost EAT, and boost term. For factoid questions, answer finding is done by using Named Entity Recognition. Answer scoring is done by calculating keyword occurrences and answer-keywords distance (MRR = 0.6191). For non-factoid questions, answer finding is done by identifying sentence pattern and clue words. Answer scoring is done by considering pattern priority and keyword occurrences (MRR = 0.8079). Keywords: monolingual, open-domain, Indonesian language, factoid, non-factoid. ∗ Copyright 2011 by Alvin Andhika Zulen and Ayu Purwarianti 1 Introduction Question Answering (QA) is a task in Natural Language Processing (NLP) that will automatically provide answers to questions posed in natural language. QA system can use a database or document collection (local or web) as the sources of the answer. A QA system usually consists of three main components (Harabagiu et al., 2000): question analyzer, passage retriever, and answer finder. Question analyzer component aims to classify the question according to the Expected Answer Type (EAT) as well as extract the keywords in question. These keywords will be used as input query in passage retriever component to get candidate documents/paragraphs that contain the answer. Answer finder component searches for candidate answers from documents/paragraphs that have been found previously. Each candidate answer will be given a score based on its compliance with the question. Some candidate answers will be selected as the best answers to the question. Each component can use various methods based on the language, question domain, question type, and available tools. QA system for Indonesian language that has been built: handle factoid questions only (Purwarianti et al., 2007; Wisudawan, 2010), handle non-factoid questions only (Yusliani, 2010), the domain of questions is limited (Mahendra et al., 2008), cross lingual and use NLP tools for English (Wijono et al., 2006; Wisudawan, 2010). There are also researches in NLP for Indonesian language that have been done, including parser, stemmer (Adriani et al., 2007), Part-of-Speech (POS) tagger (Wicaksono and Purwarianti, 2010), and Named-Entity (NE) Tagger (Budi et al., 2005). From these studies, we obtained some things that can be explored further and improved to produce a QA system for Indonesian language using monolingual approaches with wider scope 25th Pacific Asia Conference on Language, Information and Computation, pages 622–631 622 of questions that can be handled. This QA system is expected to …

Content truncated to 500 words