Learning Translations of Named-Entity Phrases from Parallel Corpora

Learning Translations of Named-Entity Phrases from Parallel Corpora Robert C. Moore Microsoft Research Redmond, WA 98052, USA bobmoore@microsoft.com Abstract We develop a new approach to learn- ing phrase translations from parallel cor- pora, and show that it performs with very high coverage and accuracy in choosing French translations of English named-entity phrases in a test corpus of software manuals. Analysis of a subset of our results suggests that the method should also perform well on more gen- eral phrase translation tasks. 1 Introduction Machine translation can benefit greatly from aug- menting knowledge of word translations with knowledge of phrase translations. Multiword phrases may have nonliteral translations, or one of several equally valid literal translations may be strongly preferred in practice. Automatically learning translations of single words from parallel corpora has been much studied over the past ten years or so (Melamed, 2000, and references), but learning translations of multiword phrases has re- ceived less attention. (See Section 5 for a review of prior work in this area.) In this paper, we de- velop a new approach to learning phrase transla- tions from parallel corpora, and show that it per- forms with very high coverage and accuracy on a named-entity phrase translation task. Moreover, analysis of a subset of our evaluation results sug- gests that the method should also perform well on more general phrase translation tasks. In our approach, we are given a sentence- aligned parallel corpus annotated with a set of phrases in one of the two languages (the source language), and our goal is identify the correspond- ing phrases in the corpus in the other language (the target language), ranking the translation pairs in order of confidence. Certain segments of the target language corpus may be annotated as constituting lexical compounds, which may or may not include the translations of the source language phrases of interest. Otherwise there is no annotation of the target language text, except for its being divided into words and sentences. Below we describe the issues in named-entity phrase translation motivating this research, we ex- plain our algorithm, and we present the results of our evaluation on a named-entity phrase transla- tion task. We pay particular attention to the subset of the data that lacks the special characteristics of the named-entity task that we take advantage of to optimize our performance, to suggest how the algorithm might perform on more general tasks. Finally we compare our approach and results to previous work on learning phrase translations. 2 The Named-Entity Phrase Translation Task Named-entity expressions (Chinchor and Marsh, 1997) are any words or phrases that name a spe- cific entity. While often thought of in terms of categories such as persons, organizations, or lo- cations, in technical text a much wider range of types of entities are often named In software man- uals, for example, named-entity expressions in- 259 clude names of menu items, dialogue boxes, soft- ware systems, etc. While named-entity expres- sions are typically used as proper nouns, those en- countered in technical …

Content truncated to 500 words