Statistical Machine Translation Improvement based on Phrase Selection

Proceedings of Recent Advances in Natural Language Processing, pages 458–464, Hissar, Bulgaria, Sep 7–9 2015. STATISTICAL MACHINE TRANSLATION IMPROVEMENT BASED ON PHRASE SELECTION Cyrine Nasri SMarT, LORIA Campus Scientifique BP 139, 54500 Vandoeuvre L`es Nancy Cedex, France cyrine.nasri@loria.fr Chiraz Latiri LIPAH laboratory Faculty of Sciences of Tunis University of Tunis El Manar Tunisia chiraz.latiri@gnet.tn Kamel Sma¨ıli SMarT, LORIA Campus Scientifique BP 139, 54500 Vandoeuvre L`es Nancy Cedex, France smaili@loria.fr Abstract This paper describes the importance of in- troducing a phrase-based language model in the process of machine translation. In fact, nowadays SMT are based on phrases for translation but their language models are based on classical ngrams. In this pa- per we introduce a phrase-based language model (PBLM) in the decoding process to try to match the phrases of a translation table with those predicted by a language model. Furthermore, we propose a new way to retrieve phrases and their corre- sponding translation by using the principle of conditional mutual information. The SMT developed will be compared to the baseline one in terms of BLEU, TER and METEOR. The experimental results show that the introduction of PBLM in the translation decoding improve the results. 1 INTRODUCTION Language modeling is a crucial task in many areas of natural language processing (NLP) like Automatic Speech Recognition (ASR), Statistical Machine Translation (SMT), Optical Character Recognition (OCR), etc. Every improvement in the language model performance can impact the previously cited applications Many researches on language modeling have been proposed in the literature over the past decades (Yoshua et al., 2003), (Schwenk, 2007) and (Wu et al., 2012). Nowadays, the new language models are based on deep learning techniques (Arsoy et al., 2012). Some studies were proposed to improve the language model quality by adding external informations (syntactic, morphological, etc). Significant improvements were noted (Char- niak et al., 2003) (Kirchhoff and Yang, 2005) (Sarikaya and Deng, 2007) (L. Schwartz and et al., 2011), (Xiao et al., 2011). In the following, we will be interested by variable- length models. In fact, words are commonly used as the basic lexical unit in standard language model, however in automatic speech recognition, some works were based on variable-length models where the basic unit is variable in terms of length. These variable-length ngrams correspond phrases as defined in the speech recognition andto machine translation communities. The models shown that they reduce the perplexity of the language model and sometimes they improve the performance of the ASR (Giachin, 1995) (Dietrich, 1998) (G. Riccardi and Riley, 1997) (K.F. Ries and Waibel, 1996) (Zitouni et al., 2003). In SMT, (Baisa, 2011), first proposed the chunk- based language model (including phrase-based) in machine translation but did not give a solution. Recently, (Xu and Chen., 2015) designed a direct algorithm for phrase-based language model in statistical machine translation. In their method, phrase can be any word sequence. The phrase vocabulary is huge and the data sparsity problem is very serious. It leads to difficulty in probability estimation for phrase-based language model. Language model is considered …

Content truncated to 500 words