Answer Extraction by Flexible Matching, Filtering, and Interpretation

Abstract

This paper describes a Korean Question Answering (KorQuA) system. In our observation, QA systems need to consider these things to be more confident and intelligent: Terms in a text should be classified as the characteristics of matching. For example, range expression doesn’t need exact matching and proper noun such as book title needs exact term matching. To support an answer for a question, the answer should not have negative or uncertain contexts. And date expression has a tendency to be represented as a relative expression in a news article such as last month and 2 years ago. Our KorQuA system takes a natural language question as an input and produces a list of answers ranked, which consists of three components: question interpretation, flexible passage retrieval, and answer extraction. For flexible text retrieval, terms in a question are represented as five data types: date, range, core, keyword, and expected answer type. Similarity of a passage is calculated by matching terms according to these data types. Answer extraction involves three procedures: entity extraction, answer filtering and answer interpretation. To extract candidate answers corresponding to the expected answer type, we identify named entities such as person, organization, location, date, quantities, durations, and linear measures, and semantic categories. We filtered candidate answers modified by negative context, uncertain context or uncertain background domain of a document. For date type question, we interpret a relative expression as a new answer which is not represented in a text by calculating an absolute date from a relative date expression based on the written date of a document.