Cutting the Long Tail: Hybrid Language Models for Translation Style Adaptation

Proceedings of the 13th Conference of the European Chapter of the Association for Computational Linguistics, pages 439–448, Avignon, France, April 23 - 27 2012. c⃝2012 Association for Computational Linguistics Cutting the Long Tail: Hybrid Language Models for Translation Style Adaptation Arianna Bisazza and Marcello Federico Fondazione Bruno Kessler Trento, Italy {bisazza,federico}@fbk.eu Abstract In this paper, we address statistical ma- chine translation of public conference talks. Modeling the style of this genre can be very challenging given the shortage of available in-domain training data. We investigate the use of a hybrid LM, where infrequent words are mapped into classes. Hybrid LMs are used to complement word-based LMs with statistics about the language style of the talks. Extensive experiments comparing different settings of the hybrid LM are re- ported on publicly available benchmarks based on TED talks, from Arabic to English and from English to French. The proposed models show to better exploit in-domain data than conventional word-based LMs for the target language modeling component of a phrase-based statistical machine transla- tion system. 1 Introduction The translation of TED conference talks1 is an emerging task in the statistical machine transla- tion (SMT) community (Federico et al., 2011). The variety of topics covered by the speeches, as well as their specific language style, make this a very challenging problem. Fixed expressions, colloquial terms, figures of speech and other phenomena recurrent in the talks should be properly modeled to produce transla- tions that are not only fluent but that also em- ploy the right register. In this paper, we propose a language modeling technique that leverages in- domain training data for style adaptation. 1http://www.ted.com/talks Hybrid class-based LMs are trained on text where only infrequent words are mapped to Part- of-Speech (POS) classes. In this way, topic- specific words are discarded and the model fo- cuses on generic words that we assume more use- ful to characterize the language style. The factor- ization of similar expressions made possible by this mixed text representation yields a better n- gram coverage, but with a much higher discrimi- native power than POS-level LMs. Hybrid LM also differs from POS-level LM in that it uses a word-to-class mapping to determine POS tags. Consequently, it doesn’t require the de- coding overload of factored models nor the tag- ging of all parallel data used to build phrase ta- bles. A hybrid LM trained on in-domain data can thus be easily added to an existing baseline sys- tem trained on large amounts of background data. The proposed models are used in addition to standard word-based LMs, in the framework of log-linear phrase-based SMT. The remainder of this paper is organized as fol- lows. After discussing the language style adapta- tion problem, we will give an overview of relevant work. In the following sections we will describe in detail hybrid LM and its possible variants. Fi- nally, we will present an empirical analysis of the proposed technique, including intrinsic evaluation and SMT experiments. 2 Background Our working scenario is the translation of TED …

Content truncated to 500 words