Data selection for compact adapted SMT models

Data Selection for Compact Adapted SMT Models Shachar Mirkin shachar.mirkin@xrce.xerox.com Xerox Research Centre Europe, Meylan, France Laurent Besacier laurent.besacier@imag.fr LIG, University Of Grenoble Alps, Grenoble, France Abstract Data selection is a common technique for adapting statistical translation models for a specific domain, which has been shown to both improve translation quality and to reduce model size. Selection relies on some in-domain data, of the same domain of the texts expected to be trans- lated. Selecting the sentence-pairs that are most similar to the in-domain data from a pool of parallel texts has been shown to be effective; yet, this approach holds the risk of resulting in a limited coverage, when necessary n-grams that do appear in the pool are less similar to in- domain data that is available in advance. Some methods select additional data based on the actual text that needs to be translated. While useful, this is not always a practical scenario. In this work we describe an extensive exploration of data selection techniques over Arabic to French datasets, and propose methods to address both similarity and coverage considerations while maintaining a limited model size. 1 Introduction Data selection (DS) is a key method for domain adaptation. It is commonly used in statistical machine translation (SMT) for selecting from a pool of parallel sentences a subset that is more similar to the target domain, and using it to train the translation model, the language model, or both. Data selection can also be used when more compact models are needed due to memory limitations, for instance. A wide variety of selection methods have been used over the years, where the main principle is to measure the similarity of sentences from the pool to some in- domain data, either the development or the (source side of the) test set. Such similarity is often based on information theory metrics, like perplexity, applied to either side of the training data (source or target) or – as often turned out to be more effective – to both. One problem these methods suffer from is that they “overfit” the in-domain data, since sentences with unseen n- grams are less likely to be selected. Some methods address this coverage issue by expanding the selected training data to include more instances of previously unseen or infrequent n-grams. Techniques based on information retrieval (IR) have also been widely used for data selection, especially for choosing training or tuning data based on the test set, either from parallel corpora or by mining comparable corpora to find relevant parallel phrases or sentences. This produces highly adapted models for the test set, and generally reduces the number of out-of-vocabulary (OOV) words. The limitation of such methods is that they typically rely on the availability of the text for translation before the final model is produced. While sometimes possible, this is not always the case. This work proposes an in-depth investigation of several data selection techniques for SMT adaptation. We experiment with less conventional and challenging domains for Arabic to French …

Content truncated to 500 words