英文名词短语事件指代消解方法研究

发布时间：2018-02-16 23:14

本文关键词： 事件指代消解 OntoNotes 4.0 结构化句法特征语义角色　出处：《太原理工大学》2016年硕士论文　论文类型：学位论文

【摘要】：指代消解是通过研究句子中的指代关系,来确定句中指代词所指的实体或者事件。指代消解一直是计算语言学领域的关键问题,关于它的研究成果在机器翻译、信息抽取等领域的应用前景十分广泛。之前关于指代消解的研究都集中在实体指代关系的消解上,使得实体指代消解已经取得了长足的发展,然而关于事件指代关系的处理研究才刚刚开始。由于事件区别于实体,具有实体不具备的特殊性,所以,事件指代关系的消解必须与实体指代消解分开研究,且具有较高研究价值和实践意义。根据指代词的不同可以分为代词的事件指代消解和名词短语的事件指代消解,名词短语本身包含了一定的句法及语义信息,这样的信息可以进一步提高指代消解的性能。所以,我们对名词短语的事件指代消解做了相关研究。本文首先研究了名词短语事件指代消解中正负例平衡问题。由于传统的样例生成方法会导致大量负例的产生,从而引起正负例比例失衡的情况,因此我们给出了一个正负例平衡方法,并在Onto Notes 4.0英文语料上进行了实验。其次,在计算事件语义相似度的元组(语义角色)中加入了时间和地点元素。由于事件的特殊性,时间和地点往往是区分事件的两个重要因素,所以使用施事者、受事者、时间以及地点四种语义角色生成事件指代消解系统的语义特征可以用来判断先行语候选是否与照应语表达了同一事件。再次,通过对结构化句法树进行四种不同剪裁,并将时间和地点两种语义角色加入到语义角色扩展树中,分析经过不同剪裁方式处理过的结构化特征对系统的影响。最后将平面特征、语义特征和结构化句法特征与双候选模型相结合进行了实验并做了对比分析,系统在Onto Notes 4.0英文语料上的系统性能达到了41.71%,与基准系统相比准确率提高了2.24%,系统性能提高了2.23%。
[Abstract]:Anaphora resolution is to determine the entity or event of the pronoun in a sentence by studying the anaphora relation in a sentence. Anaphoric resolution has always been a key problem in computational linguistics, and its research results are in machine translation. The application prospect of information extraction and other fields is very extensive. Previous researches on anaphoric resolution have focused on the resolution of entity anaphora relation, which has made a great progress in entity anaphora resolution. However, the research on the treatment of event reference relationship is just beginning. Because the event is different from the entity and has the particularity that the entity does not have, the resolution of the event reference relationship must be studied separately from the entity anaphora resolution. According to the differences of pronoun, it can be divided into pronoun event reference resolution and noun phrase event anaphoric resolution. Noun phrase itself contains certain syntactic and semantic information. Such information can further improve the performance of reference resolution. In this paper, we first study the balance of positive and negative cases in noun phrase event anaphora resolution. Because the traditional method of sample generation can lead to a large number of negative examples. Therefore, we give a positive and negative example balance method and experiment on the English corpus of Onto Notes 4.0. Secondly, The element of time and place is added to the tuple (semantic role) to calculate the semantic similarity of events. Because of the particularity of events, time and place are often two important factors to distinguish events, so the agents and recipients are used. The semantic features of the antecedent candidate can be used to determine whether the antecedent candidate represents the same event as the anaphora. Two semantic roles, time and place, are added to the semantic role extension tree to analyze the influence of structured features which have been processed by different tailoring methods on the system. Combining semantic features and structured syntax features with double candidate models, the system performance on Onto Notes 4.0 English corpus is 41.71, the accuracy is 2.2444 and the system performance is 2.23.
【学位授予单位】：太原理工大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：H314;TP391.1

【参考文献】