Text Segmentation Criteria for Statistical Machine Translation

For several reasons machine translation systems are today unsuited to process long texts in one shot. In particular, in statistical machine translation, heuristic search algorithms are employed whose level of approximation depends on the length of the input. Moreover, processing time can be a bottleneck with long sentences, whereas multiple text chunks can be quickly processed in parallel. Hence, in real working conditions the problem arises of how to optimally split the input text. In this work, we investigate several text segmentation criteria and verify their impact on translation performance by means of a statistical phrase-based translation system. Experiments are reported on a popular as well as difficult task, namely the translation of news agencies from Chinese-English as proposed by the NIST MT evaluation workshops. Results reveal that best performance can be achieved by taking into account both linguistic and input length constraints.