面向多层次篇章语义的机器翻译理论、方法与实现

Machine Translation has witnessed three different research stages ranging from word-based to phrase-based and syntax-based. It treats machine translation as a process of converting source structures to target ones at word, phrase and syntax levels, where the process is carried out at single sentence level. Therefore, how to utilize cross-sentence discourse knowledge and multi-level semantic information for machine translation is one of the fundamental issues of the state-of-the-art machine translation technologies. .. This proposal aims to research on the key technologies to address the fundamental issue aforementioned. The major focuses are: 1) from machine translation viewpoint, propose the theory, model and annotation guideline for the studies of multi-level semantics at word, sentence and discourse levels, and based on which construct machine translation-oriented multilingual corpus annotated with multi-level discourse semantic information; 2) based on the proposed theory and corpus, explore the multi-level discourse semantics-based model for machine translation and study the training, decoding and evaluation algorithms for the model; 3) develop the multi-level discourse semantics-based translation platform and Chinese-English bilingual translation systems; the constructed resources and the developed translation platform will be released and free available in the research community. .. The research results in this proposal are expected to fill in the research gap of multi-level discourse semantics-based machine translation and target to establish the foundation of theory and technology for the next generation of semantics-based machine translation technology. We believe the research achievements from this proposal are of great scientific significance and application value to the research and development of natural language processing, computational linguistics and Chinese information processing by advancing the state-of-the-art..

机器翻译历经了从基于词、短语到句法的三个发展阶段。它把翻译看作是一个从源语言到目标语言的词、短语和句法结构的转换过程，而这种转换过程是以单一句子为单位进行的。因此，跨句子的篇章信息和语言本身多层次语义信息的缺失成为目前机器翻译发展的主要问题。本项目拟围绕此问题展开创新性研究：1）研究面向机器翻译的多语多层次篇章语义的理论体系和表达方法，进而构建服务于机器翻译的多层次篇章语义标注的大规模多语对齐语料资源；2）基于上述理论体系和语料资源，建立面向多层次篇章语义的机器翻译模型，研究相应的训练和解码算法；3）开发和实现面向多层次篇章语义的机器翻译平台和汉英双向机器翻译系统，并面向学术界共享语料资源和翻译平台。本项目预期成果对在机器翻译研究中如何有效地利用多层次篇章语义知识，对推动自然语言处理、计算语言学和中文信息处理等相关学科的研究和发展，具有重要的科学意义和应用价值。

机器翻译历经了从基于词、短语到句法的三个发展阶段。它把翻译看作是一个从源语言到目标语言的词、短语和句法结构的转换过程，而这种转换过程是以单一句子为单位进行的。因此，跨句子的篇章信息和语言本身多层次语义信息的缺失成为目前机器翻译发展的主要问题。为解决上述问题，项目对基于语义和篇章的机器方法进行了深入研究，提出了基于篇章的神经网络机器翻译模型，即融合篇章信息的Transformer NMT模型，是学术界最早的篇章Transformer NMT模型，有效提升了篇章机器翻译的性能；为提升神经机器翻译系统的模型的语义表达能力，提出了动态多层表示融合的方法对多层神经网络中的信息进行融合，并在此基础上对深层网络语言建模方法进行研究，从提升模型容量和语义表达能力的角度对神经机器翻译系统进行改进；此外通过神经网络结构搜索的方式对语言建模方式进行探索，有效提升了模型自身的效率以及建模能力；对面向句法/语义机器翻译的中文句法/语义分析相关问题进行了深入探索，同时研究了统计机器翻译中源语言结构信息的有效使用问题，提出了基于句法骨架的机器翻译模型。提出和制定了汉语单语和中英文对照的篇章标注体系和规范，并基于此规范，标注了大规模篇章语料。开发了神经机器翻译平台，并在学术界共享。本项目达到了预期目标，的研究成果对机器翻译和自然语言处理，具有重要的科学意义和应用价值。