Language model adaptation with additional text generated by machine translation

Abstract

Statistical language modeling requires a large corpus for the application domain. When a large corpus is not available, the language model adaptation technique has often been used in the speech recognition research domain. This adaptation needs only a small corpus of the application domain (the "target corpus") and the corpus should be written in the language of the model. However, it is sometimes difficult to collect even a small corpus, especially of spoken language, due to its high cost. To address this problem, this paper proposes a novel scheme that generates a small target corpus in the language of the model by machine translation of the target corpus in another language. As information about adjacent words, which is necessary for a statistical language model, is stored in the translation knowledge, it can be extracted by machine translation and used for adaptation. Experiments showed that the language model improvement was about half of that which was obtained with a human collected corpus, and this provided some initial proof of the concept experiments.