Event-based open-domain Question Answering
Speaker:Michael Wiegand
Institution: Saarland University
Abstract:
In recent years, there has been a growing interest in Question Answering to develop methods for answering "event questions", such as "Who killed John F. Kennedy?" i.e. questions which ask for participating entities of events. These questions are thought to be well suited to be answered by means of linguistic analyses. This talk presents the results of my diploma-thesis in which I implemented a data-driven method for event-based answer extraction for a TREC-like QA scenario. Apart from an overview of my method, I will discuss the strengths and weaknesses of various structural (i.e. linguistic/knowledge-based) and non-structural (i.e. surface-based) features with regard to their suitability to find answers to event questions in a large text corpus like AQUAINT. In particular, I will critically assess the techniques which are involved in computing these features. They involve tools, such as syntactic parsers or named-entity taggers, syntactic resources, such as subcategorization lexicons, and semantic resources, such as WordNet or FrameNet. Additionally, I will present some of the crucial properties between event questions (in TREC 2005) and their corresponding answer sentences, which have been manually labelled in the AQUAINT corpus, to complement the results of my data-driven extraction algorithm.