Active learning for statistical phrase-based machine translation
NAACL '09 Proceedings of Human Language Technologies: The 2009 Annual Conference of the North American Chapter of the Association for Computational Linguistics
| Hi-index | 0.00 |
In the past few decades machine translation research has made major progress. A researcher now has access to many systems, both commercial and research, of varying levels of performance. In this thesis, we describe different methods that leverage these pre-existing systems as tools for research in machine translation and related fields. We first examine techniques for improving a translation system using additional text. The first method uses a monolingual corpus. Discrepancies are identified by translating a word list to a foreign language and back again. Entries where the original word and its double translation differ are used to learn word-level correction rules. The second method uses parallel bilingual data consisting of source language/target language sentence pairs. The source sentences are translated using a translation system, and a partial alignment is identified between the machine-translated sentences and the corresponding human-translated sentences in the target language. This alignment is used to generate phrase-level correction rules. Experimentally, both word-level and phrase-level correction rules result in improved translation performance. The learned word-level correction rules make 24,235 corrections on 20,000 Spanish to English translated sentences, with high accuracy. The learned phrase-level rules improve the translation performance (as measured by BLEU) of a French to English commercial system by 30%, and of a state of the art phrase-based system in a statistically significantly way. To train current statistical machine translation systems, bilingual examples of parallel sentences are used. Generating this data is costly, and currently feasible only in limited domains and languages. A fundamental question is whether every potential example is equally useful. We describe a ranking method for examples that scores individual sentence pairs based on the performance of translation systems trained on random subsets of the examples. When used to train a translation system, the top ranking examples result in a significantly better performing system than random selection of examples. Given these ranked examples, a model of example usefulness can potentially be learned to select the most useful unlabeled examples. Initial experiments show two previously used example features are good candidates for identifying useful examples. In the last part of this thesis we describe how automatic paraphrasing methods can be used to improve the accuracy of evaluation measures for machine translation. Given a human-generated reference sentence and a machine-generated translated sentence, we present a method that finds a paraphrase of the reference sentence that is closer in wording to the machine output than the original reference is. We show that using paraphrased reference sentences for evaluating a translation system output results in better correlation with human judgement of translation adequacy than using the original reference sentences.