Learning an Expert from Human Annotations in Statistical Machine Translation: the Case of Out-of-Vocabulary Words

Learning an Expert from Human Annotations in Statistical Machine Translation: the Case of Out-of-Vocabulary Words Wilker Aziz∗, Marc Dymetman†, Shachar Mirkin§, Lucia Specia‡, Nicola Cancedda†, Ido Dagan§ ∗University of S˜ao Paulo, †Xerox Research Centre Europe §Bar-Ilan University, ‡University of Wolverhampton will.aziz@gmail.com, {dymetman,cancedda}@xrce.xerox.com L.Specia@wlv.ac.uk, {mirkins,dagan}@cs.biu.ac.il Abstract We present a general method for incorporating an “expert” model into a Statistical Machine Transla- tion (SMT) system, in order to improve its perfor- mance on a particular “area of expertise”, and ap- ply this method to the specific task of finding ade- quate replacements for Out-of-Vocabulary (OOV) words. Candidate replacements are paraphrases and entailed phrases, obtained using monolin- gual resources. These candidate replacements are transformed into “dynamic biphrases”, generated at decoding time based on the context of each source sentence. Standard SMT features are en- hanced with a number of new features aimed at scoring translations produced by using different replacements. Active learning is used to discrimi- natively train the model parameters from human assessments of the quality of translations. The learning framework yields an SMT system which is able to deal with sentences containing OOV words but also guarantees that the performance is not degraded for input sentences without OOV words. Results of experiments on English-French translation show that this method outperforms pre- vious work addressing OOV words in terms of ac- ceptability. 1 Introduction When translating a new sentence, Statistical Ma- chine Translation (SMT) systems often encounter “Out-of-Vocabulary” (OOV) words, that is, words for which no translation is provided in the system phrase table. The problem is particularly severe when bilingual data are scarce or the text to be translated is not from the same domain as the data used to train the system. One approach consists in replacing the OOV word by a paraphrase, i.e. a word that is equiv- alent and known to the phrase-table. For instance, in the sentence “The police hit the protester”, if the source word “hit” is OOV, it could be replaced by its paraphrase “struck”. In previous work such paraphrases are learnt by “pivoting” through par- allel texts involving multiple languages (Callison- Burch et al., 2006) or on the basis of monolingual data and distributional similarity metrics (Marton et al., 2009). Mirkin et al. (2009) go beyond the use of para- phrase to incorporate the notion of an entailed phrase, that is, a word which is implied by the OOV word, but is not necessarily equivalent to it — for example, this could result in “hit” being replaced by the entailed phrase “attacked”. Both paraphrases and entailed phrases are obtained us- ing monolingual resources such as WordNet (Fell- baum, 1998). This approach results in higher coverage and human acceptability of the transla- tions produced relative to approaches based only on paraphrases. In (Mirkin et al., 2009) a replacement for the OOV word is chosen based on a score represent- ing how well it fits the context of the input sen- tence, combined with the global SMT score ob- tained after translating multiple alternative sen- tences …

Content truncated to 500 words