DEVELOPMENT OF ENGLISH-LATVIAN STATISTICAL MACHINE TRANSLATION SYSTEM: METHODS, RESOURCES AND FIRST RESULTS
Summary
This paper presents research and development of English-Latvian Statistical Machine Translation (SMT) prototypes for legal domain. Several methods have been investigated, i.e., phrase-based models and factored models. Translation quality has been evaluated using automated metrics (BLEU score) and human evaluation. In automatic evaluation the best score (46.44 BLEU points) was assigned to factored model trained on JRC Acquis corpus (version 3.0) which was also evaluated as the best from the human viewpoint. In addition, error analysis of SMT output was performed. This analysis showed that although the output of the best prototype demonstrated a reasonable quality, it had several frequent common errors, i.e., incorrect form, missing words and wrong word order. For the future, work on tree-based SMT and hybrid systems is proposed.
This work is licensed under a Creative Commons Attribution 4.0 International License.