FALCON: Federated Active Linguistic data CuratiON

Building hybrid machine translation systems by using an EBMT preprocessor to create partial translations Mikel Artetxe, Gorka Labaka, Kepa Sarasola IXA NLP Group, University of the Basque Country (UPV/EHU) {martetxe003@ikasle., gorka.labaka@, kepa.sarasola@}ehu.eus Abstract This paper presents a hybrid machine translation framework based on a pre- processor that translates fragments of the input text by using example-based ma- chine translation techniques. The pre- processor resembles a translation mem- ory with named-entity and chunk general- ization, and generates a high quality par- tial translation that is then completed by the main translation engine, which can be either rule-based (RBMT) or statisti- cal (SMT). Results are reported for both RBMT and SMT hybridization as well as the preprocessor on its own, showing the effectiveness of our approach. 1 Introduction The traditional approach to Machine Transla- tion (MT) has been rule-based (RBMT), but it has been progressively replaced by Statisti- cal Machine Translation (SMT) since the 1990s (Hutchins, 2007). Example-Based Machine Trans- lation (EBMT), the other main MT paradigm, has never attracted that much attention: even though it gives excellent results with repetitive text for which accurate matches are found in the parallel corpus, its quality quickly degrades as more gen- eralization is needed. Nevertheless, it has been ar- gued that, along with the raise of hybrid systems that try to combine multiple paradigms, EBMT can help to overcome some of the weaknesses of the other approaches (Dandapat et al., 2011)1. c⃝ 2015 The authors. This article is licensed under a Creative Commons 3.0 licence, no derivative works, attribution, CC- BY-ND. 1This paper refers to as hybridization to any combination of MT paradigms, no matter if they are integrated in a single In this paper, we propose one such system based on a multi-pass system combination: an EBMT preprocessor translates those fragments of the in- put text for which accurate matches are found in the parallel corpus, generating a high-quality par- tial translation that is then completed by the main translator, which can be either rule-based or statis- tical. The function of the EBMT preprocessor is therefore similar to that of Translation Memories (TM), with the difference that previously made translations are not reused to aid human translators but a MT engine. Needless to say, if the EBMT preprocessor was only able to reuse full sentences as traditional TM systems do at the most basic level, the quality of its partial translations would match that of humans, but its contribution would be negligible in most situations. At the same time, trying to increase the coverage by generalizing too much at the expense of translation quality, as tra- ditional EBMT systems do, would make the whole system pointless if the preprocessor is not able to outperform the main MT engine for the fragments it translates. This way, for our approach to work as intended, it is necessary to find a trade-off between coverage and translation quality. In this work, we take a preprocessor that reuses full sentences as our starting point and explore two generalization …

Content truncated to 500 words