A Machine Learning Approach to Answering Questions for Reading Comprehension Tests

A Machine Learning Approach to Answering Questions for :Reading Comprehension Tests Hwee Ton Ng Leong Hwee Teo* Jennifer Lai Pheng Kwan DSO National Laboratories 20 Science Park Drive Singapore 118230 {nhweetou, tleonghw, klaiphen}~dso.org.sg Abstract In this paper, we report results on answering questions for the reading comprehension task, using a machine learning approach. We eval- uated our approach on the Remedia data set, a common data set used in several recent pa- pers on the reading comprehension task. Our learning approach achieves accuracy competi- tive to previous approaches that rely on hand- crafted, deterministic rules and algorithms. To the best of our knowledge, this is the first work that reports that the use of a machine learning approach achieves competitive results on answering questions for reading compre- hension tests. 1 Introduction The advent of the Internet has resulted in a massive information explosion. We need to have an effective and efficient means of locat- ing just the desired information. The field of information retrieval (IR) is the traditional discipline that addresses this problem. However, most of the prior work in IR deal more with document retrieval rather than "in- formation" retrieval. This also applies to search engines on the Internet. Current search engines take a list of input words and return a ranked list of web pages that contain (or not contain) the given words. It is then left to the user to search through the returned list of web pages for the information that he needs. While finding the web pages that contain the desired information is an important first step, what an information seeker needs is often an answer to a question. That is, given a ques- tion, we want a system to return the exact answers to the question, and not just the doc- uments to allow us to further search for the * Leong Hwee Teo's current a~ation: Defence Med- ical Research Institute, Defence Science and Technol- ogy Agency, 1 Depot Road, Defence Technology Tower A, ~19-05, Singapore 109679 l:leonghw~dsl:a.gov, sg answers. The need for question answering (QA) sys- tems has prompted the initiation of the ques- tion answering track in TREC-8 (Voorhees and Tice, 2000) to address this problem. In the QA track, each participant is given a list of 200 questions, and the goal is to locate answers to these questions from a document database consisting of hundreds of thousands of documents (about two gigabytes of text). Each participant is to return a ranked list of the five best answer strings for each question, where each answer string is a string of 50 bytes (or 250 bytes) that contains an answer to the question. What, when, where, and who ques- tions that have explicit answers given in some document in the database are emphasized, but not why questions. In a related but independent effort, a group at MITRE has investigated question answer- ing in the context of the reading comprehen- sion task (Hirschman et al., 1999). The docu- ments in this task axe 115 …

Content truncated to 500 words