Research On The Key Technologies For Phrase-based Statistical Machine Translation Models

Machine translation(MT) is a key problem in the area of natural language processing(NLP).In the past twenty years,the application of statistical methods in MT gains more and more attentions and various statistical-based translation models have been proposed.Recently,phrase-based translation model(PBTM) has become one of the dominating statistical-based translation models.With the advantages of simplicity and robustness,PBTM has achieved excellent performances in many evaluations of MT,such as NIST and IWSLT.Based on PBTM,our objective is to improve the performance of phrase-based translation system.This thesis focuses on word alignment,reordering and syntax-based language models.The contributions are addressed as follows:Firstly,we proposed a novel word alignment model based on multiobjective evolutionary algorithms(MOEA).In this model,the word alignment problem is looked on as a multiobjective optimization problem and an efficient MOEA,i.e. NSGA-â…¡is chosen to perform the optimization.Our model is an unsupervised model, which could be easily extended through incorporating more objective functions.Secondly,we proposed a novel reordering models based on the reordering of source language chunks.The model is used as a preprocessing step of phrase-based translation models and could be well integrated with such models.At the same time, as a chunk-based model,syntax information could be concerned in the process of reordering while the entire parsing of the source sentence is not required.Finally,we proposed a novel concept of linked phrases,which is an extension of the concept of traditional phrases.Based on this concept,we proposed a method to incorporate the link-grammar-based language models in phrase-based translation models without changing their main frameworks.