Automatic Question-Answering for Discussion Board

1
Automatic Question-Answering for Discussion Board
Contact Information Name Erin Shaw, shaw_at_isi.edu Donghui Feng, donghui_at_isi.edu Project URL http//www.isi.edu/carte/den Project size Small (4 people) Project lifetime 2003 to 2005 Funder Lord Corporation Foundation (via Viterbi School of Engineering) QA for DEN/ISI Discussion Board Goal Automatic answering questions in posts for DEN/ISI Discussion Board Traditional QA in NLP Research Most are factoid questions Question contains only one sentence Answers can be given with a short phrase Data collections of unordered plain text files. Our Challenges Question presented in a post may include multiple sentences or even multiple paragraph Most answers may need an explanation Data structured posts/topics/forums in natural language.
Our Approach 1. Identify Most Related Posts/Docs Collect language features for each post/doc N-gram language model Adjusted by TFIDF scores Represent each post/doc with a vector Compute similarities between posts/docs 2. Summarize ranked most similar posts/topics 3. Further approaches for answer extraction Topic detection, discourse analysis, etc System Pipeline
Intelligent Systems Division Open House June 17,
2005