Translating transliterations

Abstract

Translating new entity names is important for improving performance in Natural Language Processing (NLP) applications such as Machine Translation (MT) and Cross Language Information Retrieval (CLIR). Usually, transliteration is used to obtain phonetic equivalents in a target language for a given source language word. However, transliteration across different writing systems often results in different representations for a given source language entity name. In this paper, we address the problem of automatically translating transliterated entity names that originally come from a different writing system. These entity names are often spelled differently in languages using the same writing system. We train and evaluate various models based on finite state technology and Statistical Machine Translation (SMT) for a character-based translation of the transliterated entity names. In particular, we evaluate the models for translation of Russian person names between Dutch and English, and between English and French. From our experiments, the SMT models perform best with consistent improvements compared to a baseline method of copying strings.