PROMT DeepHybrid system for WMT12 shared translation task

Proceedings of the 7th Workshop on Statistical Machine Translation, pages 345–348, Montr´eal, Canada, June 7-8, 2012. c⃝2012 Association for Computational Linguistics PROMT DeepHybrid system for WMT12 shared translation task Alexander Molchanov PROMT LLC 16A, Dobrolubova av. 197198, St. Petersburg, Russia Alexander.Molchanov@promt.ru Abstract This paper describes the PROMT submission for the WMT12 shared translation task. We participated in two language pairs: English- French and English-Spanish. The translations were made using the PROMT DeepHybrid engine, which is the first hybrid version of the PROMT system. We report on improvements over our baseline RBMT output both in terms of automatic evaluation metrics and linguistic analysis. 1 Introduction In this paper we present the PROMT DeepHy- brid submission for WMT12 shared translation task for two language pairs: English-French and English-Spanish. A common approach to create hybrid machine translation (MT) systems on the basis of rule-based machine translation (RBMT) systems is to build a statistical phrase-based post-editing (SPE) system using state-of-the-art SMT technologies (see Si- mard et al. 2007). An SPE system views the output of the RBMT system as the source language, and reference human translations as the target lan- guage. SPE systems are used to correct typical mistakes of the RBMT output and to adapt RBMT systems to specific domains. (Dugast et al. 2007) report on good results both in terms of automatic evaluation metrics and human evaluation for the SPE systems based on PORTAGE (Sadat et al. 2005) and Moses (Koehn et al. 2007). However, an SMT model in fact makes translation output less predictable in comparison with RBMT output. We propose a different approach to hybrid MT tech- nology. We developed and incorporated the SPE component into our translation system (the statisti- cal post-editing data is controlled by the PROMT hybrid translation engine). Besides, we have an internal language model (LM) component that scores the generated translation candidates. The remainder of the paper is organized as fol- lows: in section 2 we provide the detailed descrip- tion of our hybrid MT technology. In section 3 we evaluate the performance of the technology on two language pairs: English-French and English- Spanish. We gain improvements over the baseline RBMT system in terms of BLEU score on test sets. We also introduce the results of linguistic evalua- tion performed by our experts. Section 4 summa- rizes the key findings and outlines open issues for future work. 2 System description The PROMT DeepHybrid system is based on our RBMT engine. The baseline system has been augmented with several modules for hybrid train- ing and translation. The training technology is fully automated, but each step can be fulfilled and tuned separately. 2.1 Rule-based component PROMT covers 51 language pairs for 13 differ- ent source languages. Our system is traditionally classified as a ‘rule-based’ system. PROMT uses morphosyntactic analyzers to analyze the source sentence and transfer rules to translate the sentence 345 into the target language. The crucial component of our system is the PROMT bilingual dictionaries which contain up to 250K entries for each lan- guage …

Content truncated to 500 words