Stochastic Parse Tree Selection for an Existing RBMT System

Proceedings of the 6th Workshop on Statistical Machine Translation, pages 351–357, Edinburgh, Scotland, UK, July 30–31, 2011. c⃝2011 Association for Computational Linguistics Stochastic Parse Tree Selection for an Existing RBMT System Christian Federmann DFKI GmbH Language Technology Lab Saarbr¨ucken, Germany cfedermann@dfki.de Sabine Hunsicker DFKI GmbH Language Technology Lab Saarbr¨ucken, Germany sabine.hunsicker@dfki.de Abstract In this paper we describe our hybrid machine translation system with which we participated in the WMT11 shared translation task for the English→German language pair. Our system was able to outperform its RBMT baseline and turned out to be the best-scored participating system in the manual evaluation. To achieve this, we extended an existing, rule-based MT system with a module for stochastic selection of analysis parse trees that allowed to better cope with parsing errors during the system’s analysis phase. Due to the integration into the analysis phase of the RBMT engine, we are able to preserve the benefits of a rule-based translation system such as proper generation of target language text. Additionally, we used a statistical tool for terminology extraction to improve the lexicon of the RBMT system. We report results from both automated metrics and human evaluation efforts, including exam- ples which show how the proposed approach can improve machine translation quality. 1 Introduction Rule-based machine translation (RBMT) systems that employ a transfer-based translation approach, highly depend on the quality of their analysis phase as it provides the basis for its later processing phases, namely transfer and generation. Any parse failures encountered in the initial analysis phase will proliferate and cause further errors in the following phases. Very often, bad translation results can be traced back to incorrect analysis trees that have been computed for the respective input sentences. Hence- forth, any improvements that can be achieved for the analysis phase of a given RBMT system directly lead to improved translation output which makes this an interesting topic in the context of hybrid MT. In this paper we present a study how the rule- based analysis phase of a commercial RBMT system can be supplemented by a stochastic parser. The system under investigation is the rule-based engine Lucy LT. This software uses a sophisticated RBMT transfer approach with a long research history; it is explained in more detail in (Alonso and Thurmair, 2003). The output of its analysis phase is a parse forest containing a small number of tree structures. For our hybrid system we investigated if the existing rule base of the Lucy LT system chooses the best tree from the analysis forest and how the selection of this best tree out of the set of candidates can be improved by adding stochastic knowledge to the rule-based system. The remainder of this paper is structured in the following way: in Section 2 we first describe the transfer-based architecture of the rule-based Lucy LT engine, giving special focus to its analysis phase which we are trying to optimize. Afterwards, we provide details on the implementation of the stochastic selection component, the so-called “tree selector” which allows …

Content truncated to 500 words