UAlbany's ILQUA at TREC 2007.
Min Wu, Chen Song, Zhan Yu, Tomek Strzalkowski · 2007
TREC2007 QA track introduced a combined collection of 175GB BLOG data and 2.5GB newswire data. This introduced an additional challenge for an automatic QA system to processes data in different formats without sacrificing the accuracy. In ILQUA we added a data preprocessing component to filter out noisy blog data. ILQUA has been built as an IE-driven QA system; it extracts answers from documents annotated with named entity tags. Answer extraction methods applied are surface text pattern matching, n-gram proximity search and syntactic dependency matching. The answer patterns used in ILQUA are automatically summarized by a supervised learning system and represented in form of regular expressions which contain multiple question terms. In addition to surface text pattern matching, we also adopt N-gram proximity search and syntactic dependency matching. N-grams of question terms are matched around every named entity in the candidate passages and a list of named entities are extracted as answer candidate. These named entities then go through a multi-level syntactic dependency matching component until a final answer is chosen. This year, we modified the component that tackles “Other ” questions and applied different method in the two runs we submitted. One method utilized representative words and syntactic patterns, while the other method utilized representative words from TREC data and web data. Figure 1 gives an illustration of components, data flow and control flow of ILQUA. The following sections give detailed discussion of each component of the system, evaluation results, conclusion and future work.