An Empirical Study of Smoothing Techniques for Language Modeling-英文文献.pdf

To appear in Proceedings of the 34th Annual Meeting of the ACL, June 1996 An Empirical Study of Smoothing Techniques for Language Modeling Stanley F. Chen Joshua Goodman Harvard University Harvard University Aiken Computation Laboratory Aiken Computation Laboratory 33 Oxford St. 33 Oxford St. Cambridge, MA 02138 Cambridge, MA 02138 sfc@ goodman@ 6 9 9 1 Abstract a small number of methods (typically two) on a sin- n gle corpus and using a single training data size. As u We present an extensive empirical com- a result, it is currently difficult for a researcher to J parison of several smoothing techniques in intelligently choose between smoothing schemes. 0 the domain of language modeling, includ- In this work, we carry out an extensive 1 ing those described by Jelinek and Mer- empirical comparison of the most widely used cer (1980), Katz (1987), and Church and smoothing techniques, including those described 1 Gale (1991). We investigate for the first by Jelinek and Mercer (1980), Katz (1987), and v time how factors such as training data Church and Gale (1991). We carry out experiments 1 size, corpus (e.g., Brown versu