A decoder for syntax-based statistical MT
ACL '02 Proceedings of the 40th Annual Meeting on Association for Computational Linguistics
Incremental learning of transfer rules for customized machine translation
INAP'04/WLP'04 Proceedings of the 15th international conference on Applications of Declarative Programming and Knowledge Management, and 18th international conference on Workshop on Logic Programming
| Hi-index | 0.00 |
A statistical translation model is a mathematical model for the process of human-language translation. Model parameters are automatically estimated using a corpus of translation pairs. This is in contrast to conventional rule-based machine translation systems, in which lexical, syntactic, and semantic translation rules are manually crafted by language experts over several years. The idea of statistical machine translation was first seen in the late 1940's, but the computational power at that time was not sufficient. In the last decade, word-to-word statistical translation models regained researchers' interest, due to increasing computational power and growing volume of online training materials. This thesis introduces a more advanced statistical translation model that better exploits such growing resources. Most statistical translation models are based on word-to-word translations, i.e., the operation in a model works on each word independently. We present a new model that translates a syntactic parse tree into a foreign language sentence, in which the model operations work on each node of the syntactic parse tree. To obtain a syntactic parse tree, we use an existing parser developed elsewhere. This is to take advantage of using available linguistic resources in a statistical framework. By using a syntactic parser, we are able to use rich syntactic information embedded in a sentence, and we are able to model more linguistically-motivated word movements in language translations. We use a parser only for the channel input, so that our model works for translations from any linguistically resource-poor language to a resource-rich language such as English. We have developed an efficient training algorithm and an experimental decoding program for the syntax-based translation model. We demonstrate that the alignment accuracy for Japanese-English is more than 30% better in our model compared to previous word-to-word models, and demonstrate that the decoding performance is 10–40% better for Chinese-English and Arabic-English translations.