Research Of Machine Translation System For Lifelong Learning

Under the background of economic globalization and the vigorous development of the "the Belt and Road",cultural,academic,commercial and political exchanges between people from different countries,regions and languages are becoming more and more frequent.The machine translation model in the general domain scenario can not meet the needs of more diversified and deep-seated cross language communication.In order to solve the domain adaptation problem of multi-domain translation and the cost problem of multi-domain and multi-model,the thesis combines lifelong learning algorithms,conducts research on continuous learning of multi-domain translation tasks,and proposes a machine translation system for lifelong learning.In the research process of machine translation system for lifelong learning,the main work of the thesis is as follows:1.In order to prepare the system for the basic corpus and knowledge when learning new domain tasks,the thesis designs a set of process for automatic acquisition and knowledge mining of new domain corpus.The thesis uses a Fast DTW-based method to obtain high-quality sentence-level parallel corpus.For knowledge mining in corpus,the thesis presents KDwith Multi Info that integrates multi-information and KAwith MRC,to perform knowledge discovery and knowledge alignment in parallel corpus,respectively.For the problem of domain discrimination,the thesis presents DDwith Prompts with fewshot learning to make multi-domain discrimination for the growth of the domain.Experiments show that the above three methods have deep and accurate domain knowledge mining ability and domain discrimination ability of fast training and accurate classification,respectively.2.For the multi-domain translation of the system,the thesis develops lifelong learning machine translation technology based on data fusion and model fusion.The thesis uses the domain knowledge fusion and NDNMTWith Ins domain instance model training methods to do research on machine translation based on data fusion.Experiments show that the method has a high accuracy of knowledge translation and a strong learning ability in the new domain.In model fusion-based machine translation,NDNMTWith PT is proposed in combination with Prefix-Tuning method to enhance system scalability and flexibility.Experiments show that this method has great advantages in mitigating catastrophic forgetting.3.A translation system for lifelong learning is designed and implemented.The system integrates lifelong learning machine translation technology based on data fusion and model fusion,combined with knowledge discovery,domain discrimination and other characteristic technologies,and provides users with a fully functional and professional domain translation service.The system includes offline tasks and online services.In offline tasks,the system will continuously learn translation tasks in new domains,and mine and retain knowledge in this domain;in online services,users can use automatic domain identification,domain knowledge display and editing,translation with domain knowledge,and more translation results display and other services.The machine translation system for lifelong learning in this thesis has good domain translation capability,and the average BLEU value of each domain on the relevant test set can reach 27.42,and the overall knowledge translation accuracy rate reaches 96.98%.At the same time,this system also has application values such as convenient and low-cost deployment,strong domain relevance of translation results,and sustainable learning in multi-domain translation scenarios.