【作者】 朱慕华；

【导师】 朱靖波；

【作者基本信息】 东北大学 ， 计算机软件与理论， 2013， 博士

【摘要】 成分句法分析(又被称作短语结构句法分析)是自然语言处理的核心任务之一,被广泛应用于包括机器翻译、问答系统在内的多种自然语言处理任务。由于人工标注句法数据(在句法分析领域,这类数据的集合通常被称作树库)的出现,数据驱动的句法分析方法成为当前句法分析研究的主流方法。通常来说,训练数据越充分,句法分析器获得的句法分析性能越好。为了利用尽可能多的训练数据构建句法分析系统,本文研究基于多数据源的成分句法分析。本文主要考虑两种扩展训练数据的来源。其一为异构树库,由不同研究单位遵循不同标注规范所构建的句法树库。异构树库由于标注规范的差异所以不能直接合并使用。但是为了充分利用已经存在的人工标注数据,有必要研究如何能够应用异构树库。其二,本文也研究了半监督句法分析,特别关注如何利用无标注数据改进移进-归约句法分析。相对于异构树库而言,无标注数据更加容易获得而且数据规模更大,因此研究如何利用无标注数据进行句法分析具有更加现实的意义。本文的主要研究内容包括以下四个方面：本文提出了基于启发式解码的树库转换方法。该方法被分别应用于句法树中的词性转换和句法结构转换。应用该方法进行树库转换时只需要修改词性标注器和句法分析器的解码过程。首先,在目标树库上构建词性标注器(句法分析器),然后应用得到的词性标注器(句法分析器)对源树库中的句子进行词性标注(句法分析),并且在分析过程中引入源树库中的原有标注作为解码的指导信息。基于启发式解码的树库转换方法可以将源树库中的标注信息作用于整个解码过程,因而获得更好的性能。最终转换精度达到84.2%。本文提出了基于特征的树库转换方法。与基于启发式解码的方法相比,基于特征的转换方法并不将源树库中的标注作为解码时的硬约束,而是将其作为特征引入到词性标注器和句法分析器中。首先,在源树库上构建词性标注器(句法分析器)并且应用该词性标注器(句法分析器)对目标树库中的句子进行词性标注(句法分析)。如此,目标树库中的句子同时具有两种(遵循不同规范的)标注。在这个新数据上构建用于转换的词性标注器(句法分析器)对源树库进行转换。最终的转换精度达到84.8%。本文提出了基于协同解码的异构树库句法分析方法。树库转换是间接利用异构树库的方法,而协同解码属于直接利用异构树库的方法。协同解码方法首先在每一个异构树库上分别构建句法分析器,然后应用得到的多个句法分析器对测试数据进行句法分析。在分析过程中,句法分析器参照彼此的分析结果,使句法分析器的分析结果尽可能达成结构上的一致。协同解码方法在两个实验数据上分别获得0.5%和0.7%的性能提高。本文研究了半监督移进-归约句法分析,利用无标注数据改进词性标注器和移进-归约句法分析器。本文方法首先利用集成句法分析器对大规模无标注数据进行句法分析,然后从得到的自动分析树中抽取词性部分作为额外的训练数据构建词性标注器。这样得到的词性标注器可以生成更有利于句法分析的词性标注结果。另外从集成句法分析器生成的自动分析树中抽取词汇依存关系,并且在依存关系的基础上设计新的特征改进移进-归约句法分析器。最终将词性标注和句法分析两方面的工作结合在一起,可以得到目前最好的移进-归约成分句法分析结果。在英文和中文上的性能分别达到90.9%和82.2%。基于本文的技术,我们成功开发了多套性能优异的成分句法分析系统并且在机器翻译和语义角色标注等任务上得到了实际应用。更多还原

【Abstract】 Constituent parsing (also known as phrase-structure parsing) is one of core tasks of natural language processing, which often serves as an important and fundamental component in other tasks, such as machine translation and question answering. Since the release of human-labeled treebanks, data-driven approaches have become the main stream of constituent parsing. Generally, more training data results in parsers with higher parsing accuracies. Our work focuses on exploiting data from multiple sources to enlarge training data of constituent parsers. Here we consider two types of data. One is heterogeneous treebanks, which are treebanks constructed in different organizations and following different annotation standards. Taking into consideration the high cost of annotating parse trees by human, it is practically significant to make full use of treebanks that have already been created. The other type of data that we are interested in is unlabeled data. In contrast to heterogeneous treebanks, unlabeled data is relatively easy to obtain and is generally available in large scale. Our contributions are summarized as follows:We propose to apply an informed decoding approach to treebank conversion. The approach is applied to convert POS tags and syntactic structures, respectively. We fist build a POS tagger (syntactic parser) on the target treebank and then apply the POS tagger (syntactic parser) to process sentences in a source treebank. During the decoding phase of the POS tagger (syntactic parser) on the source treebank, original annotations in the source treebank are used as guiding information of the decoding. The informed-decoding approach reaches a conversion accuracy of 84.2%We propose to apply a feature-based approach to treebank conversion. In contrast to the informed-decoding approach, the feature-based approach encodes annotations in a source treebank as features instead of hard constraints. We first build a POS tagger (syntactic parser) on the source treebank and then apply the POS tagger (syntactic parser) to the sentences of the target treebank. After that, the sentences in the target treebank contain two types of annotations, based on which we build a new POS tagger (syntactic parser) that is used to conduct conversion. The feature-based approach improves the conversion accuracy to 84.8%.We propose to do heterogeneous parsing through collaborative decoding. Compared to treebank conversion, heterogeneous parsing via collaborative decoding is a direct way to use heterogeneous treebanks. The idea of collaborative decoding is to build a parser on individual treebanks, and then apply the resulting parsers to parse testing sentences simultaneously. During the decoding phase, consensus information between the decoders is incorporated to encourage parsers to reach consensus in parsing results. On the two experimental datasets, the co-decoding approach achieves an improvement of 0.5% and 0.7%, respectively.We study semi-supervised shift-reduce constituent parsing. The basic idea is to use an integrated parser to process unlabeled data to obtain a large set of auto-parsed trees. The POS data extracted from the auto-parsed trees is used as additional data to train stand-alone POS taggers, which can provide syntactic parsers with better POS tags. We also extract reliable partial information from auto-parsed trees. Specifically, we use lexical dependency information, based on which we design a set of novel features. Combining the improved stand-alone POS taggers and improved shift-reduce parsers, we finally advance shift-reduce parsing to the state-of-the-art. The resulting parser reaches the accuracy of 90.9% and 82.2% on English and Chinese, respectively.Based on the technologies discussed in the thesis, we developed several state-of-the-art syntactic parsers, which have been deployed successfully in natural language processing systems like machine translation and semantic role labeling.更多还原