Unsupervised training of maximum-entropy models for lexical selection i in rule-based machine translation

Unsupervised training of maximum-entropy models for lexical selection in rule-based machine translation Francis M. Tyers HSL-fakultehta, UiT Norgga ´arktalaˇs universitehta, N-9018 Romsa Felipe S´anchez-Mart´ınez Dept. Lleng. i Sist. Inform., Universitat d’Alacant, E-03071 Alacant Mikel L. Forcada Dept. Lleng. i Sist. Inform., Universitat d’Alacant, E-03071 Alacant Abstract This article presents a method of training maximum-entropy models to perform lexi- cal selection in a rule-based machine trans- lation system. The training method de- scribed is unsupervised; that is, it does not require any annotated corpus. The method uses source-language monolingual corpora, the machine translation (MT) system in which the models are integrated, and a sta- tistical target-language model. Using the MT system, the sentences in the source- language corpus are translated in all possi- ble ways according to the different transla- tion equivalents in the bilingual dictionary of the system. These translations are then scored on the target-language model and the scores are normalised to provide frac- tional counts for training source-language maximum-entropy lexical-selection mod- els. We show that these models can per- form equally well, or better, than using the target-language model directly for lexical selection, at a substantially reduced compu- tational cost. 1 Introduction Corpus-based machine translation (MT) has been the primary research direction in the field of MT in recent years. However, rule-based MT (RBMT) systems are still being developed, and there are many successful commercial and non-commercial systems. One reason for the continued development of RBMT systems is that in order to be successful, c⃝ 2015 The authors. This article is licensed under a Creative Commons 3.0 licence, no derivative works, attribution, CC- BY-ND. corpus-based MT requires parallel corpora in the order of tens of millions of words. Although for some language pairs these exist, they only exist for a fraction of the world’s languages. An RBMT system typically consists of an analy- sis component,1 a transfer component and a genera- tion component. As part of the transfer component it is necessary to make choices regarding words in the source language (SL) which may have more than one translation in the target language (TL). Lexical selection is the task of choosing, for a given SL word, the most adequate translation in the TL among a known set of alternatives. The task is related to the task of word-sense disambiguation (Ide and V´eronis, 1998). However, it is different to word-sense disambiguation in that lexical selection is a bilingual problem, not a monolingual problem: its aim is to find the most adequate translation, not the most adequate sense. Thus, it is not necessary to choose among a series of fine-grained senses if all these senses result in the same final translation; however, it may sometimes be necessary to choose a different translation for the same sense, for example in a collocation. 1.1 Prior work Dagan and Itai (1994) used the term word sense dis- ambiguation to refer to what is actually lexical se- lection in MT; they used a parser to identify syntac- tic relations such as subject–object or subject–verb. …

Content truncated to 500 words