【作者】 马建军；

【导师】 黄德根； 高庆狮；

【作者基本信息】 大连理工大学 ， 计算机应用技术， 2012， 博士

【摘要】 英语名词短语识别在机器翻译中有着重要的作用。现有英汉机器翻译的瓶颈之一就是名词短语的歧义消解问题。本文提出一种英语功能短语及其自动识别方法,以消除英汉机器翻译中的名词短语结构歧义。基于名词短语在小句中的功能语法来确定名词短语的边界,这样的名词短语在本文中称为功能名词短语。研究工作主要包括英汉机器翻译中的名词短语结构歧义问题分析、英文词性自动标注和英文短语识别等研究。本文构建了一个小规模的商务领域英汉双语平行语料库作为研究语料,包含20万英文词和27万汉字。(1)英汉机器翻译中的名词短语结构歧义问题。基于忠实度与流畅度合并的机器翻译人功能评价方法,对SYSTRAN和GOOGLE机译系统的英汉机器翻译结果进行评测,分析了机器翻译中的词义歧义和句法歧义问题；在此基础上,归纳了机器翻译中名词短语的结构歧义问题。研究表明,不论是词义歧义问题还是句法歧义问题,都与名词短语的识别和理解有很大的关系。其中,N1+prep+N2是引起歧义的最典型的表层结构,这种表层结构容易引起的歧义主要体现在：与动词构成固定搭配的名词歧义问题；小品词引起的歧义问题；“介词+名词”作后置定语的歧义问题；以及“介词+名词”作状语的歧义问题等四种名词短语结构歧义。(2)研究面向英汉机器翻译的英语词性标注。研究了一个应用于机器翻译的英语词性标注系统,为功能名词短语的识别研究提供词性知识。在预实验的基础上,对宾州树库标注集进行细化和改进,形成了本文的词性标注集。依据此标注集,采用最大熵模型结合语言规则的方法进行了标注词性。实验结果表明,开放测试的精确率达到98.14%,未登录词的精确率为85.65%。(3)研究英语功能名词短语识别。本文识别了名词短语的边界和句法功能。首先依据系统功能语法,归纳了功能名词短语在小句中的功能,形成了本文的功能块标注集；其次选择了条件随机域模型结合语义信息和规则的方法进行了名词短语识别。为检验本文的词性标注集在功能名词短语识别中的作用,在开放测试中还选择了斯坦福标注器作为比较。实验结果表明,结合金标准词性实验的F值达到了89.04%。此外,使用本文的词性标注集有助于提高名词短语的识别,比使用宾州树库标注集提高了2.21%。更多还原

【Abstract】 English NP chunking plays an important role in machine translation. One major problem with machine translation lies in its ability to resolve the ambious problems caused by nouns. This paper, therefore, presents a study on the automatic identification of a kind of English functional noun phrases (NP), on the purpose of resolving structural ambiguity caused by noun phrases in English-Chinese machine translation (MT). Functional noun phrases refer to those noun phrases which are defined based on their syntactic functions in clauses. Structural ambiguity caused by noun phrases then can be solved by identifying their syntactic functions. The study includes the following three aspects:the analysis of the ambiguity problems in English-Chinese machine translation, the MT-oriented English Part-of-speech (POS) tagging, and the NP chunking. This NP chunking study is made on a self-built English-Chinese parallel corpus of in business domain which consists of200,000English words and270,000Chinese characters. The main research work can be summarized as follows:(1) The analysis of the structural ambiguity problems caused by noun phrases in English-Chinese machine translation. This paper makes a comparative analysis of the ambiguity resolution of two MT approaches:Rule Based Machine Translation (RBMT) and Statistical Machine Translation (SMT), by comparatively analyzing the Chinese translation work translated by SYSTRAN and GOOGLE translation systems, based on a manual machine translation evaluation method combining both faithfulness and smoothness. The results show that both lexical ambiguity and structural ambiguity have a lot do with NP chunking and understanding. And a surface structure N1+prep+N2is a typical structure which has caused ambiguity problem. Four main structural ambiguity problem caused by NPs are ambiguity caused by NPs when they make an inseparate part of the verb phrase, ambiguity caused by particles, ambiguity caused by "prep+noun" structures when they function as postmodifers, and ambiguity caused by "prep+noun" structures when they function as adjuncts.(2) The MT-oriented English POS tagging. This tagger is supposed to provide the following NP chunking task with POS tags, for the purposed of machine translation. After the result analysis of a pre-test, a new tagset is made for MT purpose, which is based on the Penn Treebank tag set, and the English sentences in the corpus are annotated on this new tagset. The statistical method combining the rules is applied in the study; the maximum entropy model is adopted, and rule-based approach is used in post-processing. Experiments show that our tagger achieves an accuracy of98.14%in open test, and85.65%correct on unknown words.(3) The NP chunking. Both the scope of NP chunks and the syntactic function types of NP chunks are identified in this task. The function tags of noun phrases are categorized, based on the systematic functional grammar. The conditional random fields model is adopted combining both the semantic information and language rules. The system performance is further compared with that of the model trained using Stanford POS tags in the open tests. Test results show that the system has achieved an F-score of89.04%in the open test using our gold standard tags, which also proves that our new tagset is a better approach for NP chunking, which has increased the F-score by2.21%, compared with the model using the Penn Tree bank POS tags.更多还原