Using singular value decomposition to compute answer similarity in a language independent approach to question answering

Authors:
Edward Whittaker;Josef Novak;Matthias Heie;Shuichiro Imai;Sadaoki Furui
Affiliations:
Tokyo Institute of Technology, Tokyo, Japan;Tokyo Institute of Technology, Tokyo, Japan;Tokyo Institute of Technology, Tokyo, Japan;Tokyo Institute of Technology, Tokyo, Japan;Tokyo Institute of Technology, Tokyo, Japan
Venue:
LKR'08 Proceedings of the 3rd international conference on Large-scale knowledge resources: construction and application
Year:
2008

Citing 3
Cited 1

Learning human-like knowledge by singular value decomposition: a progress report

NIPS '97 Proceedings of the 1997 conference on Advances in neural information processing systems 10
Monolingual web-based factoid question answering in Chinese, Swedish, English and Japanese

MLQA '06 Proceedings of the Workshop on Multilingual Question Answering
Probabilistic latent semantic analysis

UAI'99 Proceedings of the Fifteenth conference on Uncertainty in artificial intelligence

Question answering using statistical language modelling

Computer Speech and Language

Quantified Score

Hi-index	0.00

Visualization

Abstract

In this paper we report on new developments in our datadriven, non-linguistic, language-independent approach to Question Answering (QA). In particular, we describe a new implementation of the filter-model, which is used for answer typing, where we employ the Singular Value Decomposition (SVD) in a variation on the popular Latent Semantic Analysis technique. We also describe refinements to the open-source SVD code that we used which enable us to perform the SVD on arbitrarily large matrices. Finally, we discuss results from the TREC 2005 and TREC 2006 QA evaluations in which we applied these new techniques, and compare them to results achieved with our previous filter-model approach. In particular, we show that our new filter-model using the SVD achieves an average absolute gain of around 8% and an average relative gain of nearly 59% over our previous approach for top one answer accuracy. By using both approaches in combination we are able to increase the absolute gain to approximately 10% and the relative gain to 67%.