Predicting Answer Location Using Shallow Semantic Analogical Reasoning in a Factoid Question Answering System

Copyright 2012 by Hapnes Toba, Mirna Adriani, and Hisar Maruli Manurung 26th Pacific Asia Conference on Language,Information and Computation pages 246–253 Predicting Answer Location Using Shallow Semantic Analogical Reasoning in a Factoid Question Answering System Hapnes Toba, Mirna Adriani, and Ruli Manurung Faculty of Computer Science Universitas Indonesia Depok 16424, Indonesia hapnes.toba@ui.ac.id, {mirna, maruli}@cs.ui.ac.id Abstract In this paper we report our work on a factoid question answering task that avoids named- entity recognition tool in the answer selection process. We use semantic analogical reason- ing to find the location of the final answer from a textual passage.We demonstrate that without employing any linguistic tools during the an- swer selection process, our approach achieves a better accuracy than a typical factoid ques- tion answering architecture. 1 Introduction The task of a question answering system (QAS) is to provide a single answer for a given natural language question. In a factoid QAS, the system tries to give the best answer of an open-domain fact-based ques- tion. For example, the question “Where was an Ovi- raptor fossil sitting on a nest discovered?”. A QAS should return ‘Mongolia’s Gobi Desert’ as the final answer. A typical pipeline architecture in a fact-based QAS consists of four main processes, i.e.: question analysis, query formulation, information retrieval and answer selection. The main source of complexi- ty in a QAS lies in the question analysis and answer selection process rather than in the information re- trieval (IR) phase, which is usually achieved by uti- lizing third-party modules such as Lucene, Indri, or a web search engine. The question analysis process seeks to determine the type of a given question, which in turn provides the expected answer type (EAT) of that question as a specific fact type, such as person, organization or location. The EAT will be used to select the best answer during the answer selection process, usually by utilizing a named-entity recognizer (NER) tool in a factoid QAS (Schlaefer et al., 2006). Different approaches have been used in order to improve the performance of the answer selection component. Ko et al. (2010) employed probabilistic models for an- swer ranking of NER-based answer selection by uti- lizing external semantic resources such as WordNet. More advanced techniques utilizing linguistic tools have been proposed in Sun et al. (2005), which uses syntactic relation analysis to extract the final answer, and Moreda et al. (2010), which employs semantic roles to improve NER-based answer selection. Re- cent work by Moschitti and Quarteroni (2011) pro- posed classification of paired texts that learn to s- elect answers by applying syntactic tree kernels to pairs of questions and answers. In our current work, we try to reduce the depen- dency of the answer selection process on linguistic tools such as NER systems. Our main concern is that in reality we do not always have a complete N- ER tool for every fact type. In our example men- tioned above, the answer has a fact type which is neither an exact location, person nor an organiza- …

Content truncated to 500 words