Tom Vanallemeersch

Language model adaptation for ASR of spoken translations using phrase-based translation models and named entity models

Language model adaptation based on Machine Translation (MT) is a recently proposed approach to improve the Automatic Speech Recognition (ASR) of spoken translations that does not suffer from a common problem in approaches based on rescoring i.e. errors made during recognition cannot be recovered by the MT system. In previous work we presented an efficient implementation for MT-based language model adaptation using a word-based translation model. By omitting renormalization and employing weighted updates, the implementation exhibited virtually no adaptation overhead, enabling its use in a real-time setting. In this paper we investigate whether we can improve recognition accuracy without sacrificing the achieved efficiency. More precisely, we investigate the effect of both state-of-the-art phrase-based translation models and named entity probability estimation. We report relative WER reductions of 6.2% over a word-based LM adaptation technique and 25.3% over an unadapted 3-gram baseline on an English-to-Dutch dataset.

Intellingo: An Intelligible Translation Environment

Translation environments offer various translation aids to support professional translators. However, translation aids typically provide only limited justification for the translation suggestions they propose. In this paper we present Intellingo, a translation environment that explores intelligibility for translation aids, to enable more sensible usage of translation suggestions. We performed a comparative study between an intelligible version and a non-intelligible version of Intellingo. The results show that although adding intelligibility does not necessarily result in significant changes to the user experience, translators can better assess translation suggestions without a negative impact on their performance. Intelligibility is preferred by translators when the additional information it conveys benefits the translation process and when this information is not part of the translators readily available knowledge.

Semantics-based pretranslation for SMT using fuzzy matches

Semantic knowledge has been adopted recently for SMT preprocessing, decoding and evaluation, in order to be able to compare sentences based on their meaning rather than on mere lexical and syntactic similarity. Little attention has been paid to semantic knowledge in the context of integrating fuzzy matches from a translation memory with SMT. We present work in progress which focuses on semantics-based pretranslation before decoding in SMT. This involves applying fuzzy matching metrics based on lexical semantics and semantic roles, aligning parse trees based on semantic roles, and pretranslating matching source sentence parts using aligned tree nodes.

Automated detection of inconsistent phraseology translation

Abstract We investigate the extent to which the detection of phraseological (in)consistency in the translation process can be automated. We describe the acquisition of a large corpus of Belgian legal documents consisting of French arrests translated into Dutch. We apply the sentence alignment tool GMA to the corpus, and extract phraseological unit candidates from the sentence pairs through the term candidate extraction tool TermCalc and the word alignment data produced by the GIZA++ tool. The candidates are compared to a reference set from a manual study of an MA student at Lessius/ KULeuven. They appear to cover only 33% of the bilingual phraseological unit pairs and only four French units with more than one Dutch equivalent. This indicates the need for devising techniques specifically aimed at detecting multiple equivalence, hence potential phraseological inconsistency.