Latent Morpho-Semantic Analysis: Multilingual Information Retrieval with Character N-Grams and Mutual Information

Proceedings of the 22nd International Conference on Computational Linguistics (Coling 2008), pages 129–136 Manchester, August 2008 Latent Morpho-Semantic Analysis: Multilingual Information Retrieval with Character N-Grams and Mutual Information Peter A. Chew, Brett W. Bader Sandia National Laboratories P. O. Box 5800 Albuquerque, NM 87185, USA {pchew,bwbader}@sandia.gov Ahmed Abdelali New Mexico State University P.O. Box 30002, Mail Stop 3CRL Las Cruces, NM 88003-8001, USA ahmed@crl.nmsu.edu Abstract We describe an entirely statistics-based, unsupervised, and language- independent approach to multilingual information retrieval, which we call La- tent Morpho-Semantic Analysis (LMSA). LMSA overcomes some of the shortcomings of related previous ap- proaches such as Latent Semantic Analysis (LSA). LMSA has an impor- tant theoretical advantage over LSA: it combines well-known techniques in a novel way to break the terms of LSA down into units which correspond more closely to morphemes. Thus, it has a particular appeal for use with morpho- logically complex languages such as Arabic. We show through empirical re- sults that the theoretical advantages of LMSA can translate into significant gains in precision in multilingual infor- mation retrieval tests. These gains are not matched either when a standard stemmer is used with LSA, or when terms are indiscriminately broken down into n-grams. 1 Introduction As the linguistic diversity of textual resources increases, and need for access to those resources grows, there is also greater demand for efficient © 2008. Licensed under the Creative Commons Attribution- Noncommercial-Share Alike 3.0 Unported license (http://creativecommons.org/licenses/by-nc-sa/3.0/). Some rights reserved. information retrieval (IR) methods which are truly language-independent. In the ideal but pos- sibly unattainable case, an IR algorithm would produce equally reliable results for any language pair: for example, a query in English would re- trieve equally good results in Arabic as in French. A number of developments in recent years have brought that goal more within reach. One of the factors that severely hampered early attempts at machine translation, for example, was the lack of available computing power. However, Moore’s Law, the driving force of change in computing since then, has opened the way for recent progress in the field, such as Statistical Machine Translation (SMT) (Koehn et al. 2003). Even more closely related to the topic of the pre- sent paper, implementations of the Singular Value Decomposition (SVD) (which is at the heart of LSA), and related algorithms such as PARAFAC2 (Harshman 1972), have become both more widely available and more powerful. SVD, for example, is available in both commer- cial off-the-shelf packages and at least one open- source implementation designed to run on a par- allel cluster (Heroux et al. 2005). Despite these advances, there are (as yet) not fully surmounted obstacles to working with cer- tain language pairs, particularly when the lan- guages are not closely related. This is demonstrated in Chew and Abdelali (2008). At least in part, this has to do with the lexical statis- tics of the languages concerned. For example, because Arabic has a much richer morphological structure than English and French (meaning is varied through the addition of …

Content truncated to 500 words