An Evaluation of N-Gram Correspondence Models for Transliteration Detection

Abstract

Transliteration detection (TD) is a natural language processing (NLP) subtask that is used to find matching Named Entities (NEs) from parallel or comparable text where each language text is in a different writing system. The task is aimed at building high quality transliteration lexicons for improving performance in cross language applications like machine translation (MT) and cross language information retrieval (CLIR). Recent evaluations of TD methods (for example those from the NEWS 2010 transliteration mining shared task (Kumaran et al., 2010)) underscore the need for more methods that can improve TD performance. This paper contributes to this need by evaluating the use of source-target language n-gram correspondences for TD. We present TD experiments that use three different classes of n-gram correspondence models on standard transliteration datasets from the 2009 and 2010 shared tasks on transliteration generation (Li et al. (Report of NEWS 2009 machine transliteration shared task. Proceedings of the 2009 Named Entities Workshop: Shared Task on Transliteration, 2009); Li et al. (Report of NEWS 2010 transliteration generation shared task. Proceedings of the 2010 Named Entities Workshop, 2010)). Results show relatively significant TD performance improvements between the use of lower order and higher order n-gram correspondence models, and between the different classes of n-gram correspondence models. We show that our TD experimental setup is more complex than that in related work (Li et al. (Report of NEWS 2010 transliteration generation shared task. Proceedings of the 2010 Named Entities Workshop, 2010)) in terms of the search effort required, and that the best TD performances from the n-gram correspondence models in this paper are comparable to those from state-of-the-art methods. Our results also show that the n-gram size that should be used for developing high quality transliteration models in different languages and writing systems varies. Our work therefore serves to provide preliminary insight to the n-gram sizes required to model related high quality TD models for different language pairs and writing systems.