Question-answering (QA) systems have been around since the late 1960s, but they are still not widely used. There are many reasons for this. One is the challenge of submitting only questions for which answers exist in the document collection--otherwise, the user, being unaware that the database contains answers to a limited number of questions, receives no answer, is disappointed, and becomes discouraged about using the QA system. Another reason is the limitation of acceptable question syntax; users are typically unaware of this and only obtain a correct response from the system when they make a lucky guess. In order to avoid these scenarios, some QA systems present a list of the available questions in terms of both format and content. Another issue is the level of expert knowledge required from the user; those who do not know the jargon of the domain cannot really benefit from QA systems that require the use of specific terminology. Most information extraction techniques rely on statistical models that are more or less similar to the one proposed in this paper.
Ko et al. describe a framework for answer selection in English-Japanese and Chinese monolingual and bilingual QA systems. Their experiments are based on a system that was part of the 2005 National Institute of Informatics (NII) Test Collection for Information Resources (NTCIR). According to reports from the Text Retrieval Conference (TREC), Cross-Language Evaluation Forum (CLEF), and NTCIR campaigns, QA systems’ performance reaches at best 70 percent for monolingual and 40 percent for multilingual systems. Regarding QA systems’ typical architecture, the authors note that the filtering component is the crucial point where most systems’ efficiency is significantly reduced.
This paper proposes an approach to ranking candidate answers based on logistic regression and independent graphical joint prediction models that explore the relevance and similarity of each candidate to the correct answer. The methodology is described in detail and should be of interest to developers of Asian-language QA systems and students working in natural language processing. However, the solution offered by the authors does not seem innovative, based on other state-of-the-art developments in the field.