Studying Feature Generation from Various Data Representations forAnswer Extraction

摘要

In this paper, we study how to generatefeatures from various data representations,such as surface texts and parse trees, foranswer extraction. Besides the featuresgenerated from the surface texts, wemainly discuss the feature generation inthe parse trees. We propose and comparethree methods, including feature vector,string kernel and tree kernel, to representthe syntactic features in Support VectorMachines. The experiment on the TRECquestion answering task shows that thefeatures generated from the more structureddata representations significantlyimprove the performance based on thefeatures generated from the surface texts.Furthermore, the contribution of the individualfeature will be discussed in detail.