A probabilistic feature-based fill-up for SMT

A Probabilistic Feature-Based Fill-up for SMT Jian Zhang zhangj@computing.dcu.ie Liangyou Li liangyouli@computing.dcu.ie Andy Way away@computing.dcu.ie Qun Liu qliu@computing.dcu.ie The CNGL Centre for Global Intelligent Content,School of Computing,Dublin City University, Ireland Abstract In this paper, we describe an effective translation model combination approach based on the estimation of a probabilistic Support Vector Machine (SVM). We collect domain knowledge from both in-domain and general-domain corpora inspired by a commonly used data selection algorithm, which we then use as features for the SVM training. Drawing on previous work on binary-featured phrase table fill-up (Nakov, 2008; Bisazza et al., 2011), we substitute the binary feature in the original work with our probabilistic domain-likeness feature. Later, we design two experiments to evaluate the proposed probabilistic feature-based approach on the French-to-English language pair using data provided at WMT07, WMT13 and IWLST11 trans- lation tasks. Our experiments demonstrate that translation performance can gain significant improvements of up to +0.36 and +0.82 BLEU scores by using our probabilistic feature-based translation model fill-up approach compared with the binary featured fill-up approach in both experiments. 1 Introduction Like many machine-learning problems, Statistical Machine Translation (SMT) is a data- dependent learning approach. The prerequisite is large amounts of training data in order to generate statistical models. In general, the training data has to be sentence-aligned and bilin- gual. Some heuristic approaches are often used when deconstructing the training data into phrase-level representations, and the statistical models are computed based on the phrase prob- ability distributions. The generated models are then combined in a log-linear model (Och and Ney, 2002). A basic SMT system may consist of a translation model and a language model, where the translation model provides a target-language translation e for a source-language sen- tence f, and the language model ensures the fluency of the target-language translation e. One challenge which rises above others in SMT is that the translation performance de- creases when there are dissimilarities between the training and the testing environments. This type of challenge is often defined as “domain adaptation” in previous work. The underlying reasons that caused domain adaptation challenge are many, but the obvious one is that SMT system training is a complicated data-dependent processing pipeline. It often involves many efforts from various steps, for example, the phrase pair extraction step needs to be consistent with the preceding word alignment step, with one assumption made being that the context of the extracted phrases is irrelevant. In addition, the training sentences are only implicitly visible and become unnecessary once the phrase table is built. In this paper, we try to address the problem Al-Onaizan & Simard (Eds.) Proceedings of AMTA 2014, vol. 1: MT Researchers Vancouver, BC © The Authors 96 of phrase-table extraction in a phrase-based SMT training environment, and propose a proba- bilistic feature-based translation model fill-up approach by creating an inheritance relationship between the extracted phrase pairs and the corresponding bilingual sentence pairs. Domain adaptation for SMT is a well studied research field. Recently, many new ideas have been introduced, mainly …

Content truncated to 500 words