practical statistical machine translation

点击翻译按钮获取中文摘要

Currently, the dominant Statistical Machine Translation (SMT) paradigm capitalizes on pairs of sequences of words, namely phrases. Despite its success, phrase-based SMT has several well-understood shortcomings among which: no (or little) syntactical or semantic sensitivity, bad generalization capabilities, and a strong dependency on large parallel data. This severely limits in practice the deployment of SMT as a standalone technology. In this NSERC Research Program, I want to tackle these limitations by investigating three avenues.In the first theme, I will investigate ways of building translation engines based on dependency n-grams: a head word (called governor) and the set of its dependents in a sentence, as identified by a dependency parser (DP) or a comparable device. DP-ns have several advantages over standard phrases including a better compromise between lexical expressiveness and generalization power.The second theme of this research proposal is intended to solve part of a key challenge in the MT community: demonstrating that semantic information benefits SMT. As intuitive as this might appear, we have yet to demonstrate this empirically.The third and last avenue pursued in this proposal is the one of alleviating the strong dependency current SMT technology has on the availability of parallel training material. Parallel corpus enrichment as well as improved parallel data mining will be investigated.