Pattern Learning and Knowledge Annotation for Question Answering

Question answering – a type of information retrieval that deals with natural language questions – is a task that has been adressed by numerous systems following a variety of linguistic and statistical approaches. In this project, I developed the Ephyra question answering engine, which aims to be a modular and extensible framework that allows to integrate multiple approaches to question answering in one system. The current Ephyra system follows a pattern-learning approach. It automatically learns text patterns that can be applied to text passages for answer extraction. The system can be trained on question-answer pairs, using conventional web search engines to fetch passages suitable for pattern extraction. I propose a new approach to question interpretation that abstracts from the original formulation of the question, which helps to improve both precision and recall of answer extraction. In addition, Ephyra deploys knowledge annotation to support frequent question classes and questions that are hard to address with generic methods (e.g. definition questions, queries for the weather). Answers to such questions are extracted directly from structured web sites or web services.