【作者】 赵妍妍；

【导师】 秦兵；

【作者基本信息】 哈尔滨工业大学 ， 计算机科学与技术， 2007， 硕士

【摘要】 事件抽取是信息抽取领域一个重要的研究方向。事件抽取主要把人们感兴趣的,用自然语言表达的事件以结构化的形式呈现出来,如什么人,什么地方,什么时间,做了什么事等,在自动文摘,自动问答以及信息检索等领域有着广泛的应用。本文主要针对中文事件抽取中的两大主要任务:事件类别识别以及事件元素识别进行了深入研究,并实现了中文事件抽取系统HIT-IR EES。在事件类别识别中,由于语料的限制训练实例较少,存在着严重的数据稀疏问题,如何解决数据稀疏问题以及发现新事件成为研究的重点。本文提出了一种基于自动扩展事件触发词的方法来解决这些问题,首先使用辞典自动对训练集中的事件触发词进行扩展,根据扩展的触发词获取候选事件及其唯一的候选类别;然后结合扩展的触发词,使用词法、上下文信息和辞典特征,利用最大熵分类器对候选事件进行二元分类,来判定事件的类别。该方法将候选事件作为训练实例,有效缩减了训练实例中反例的个数,解决了训练数据正反例不平衡问题,并且扩展的触发词有助于新事件的发现;二元分类避免了多元分类带来的准确率低的问题。在ACE2005语料上的实验结果显示,最终的F值达到了61.24%,超过了传统的基于机器学习的方法。在事件元素识别中,如何从众多的Entity、Time Expression以及Value中找出事件元素,并准确的进行角色标注成为研究的重点。本文提出了两种方法:基于SRL(语义角色标注)的事件元素识别,以及基于最大熵的事件元素识别。基于SRL的方法直接利用SRL技术模块的结果,并通过角色标签转换完成事件元素识别的任务,但由于该方法过度依赖于底层技术,如SRL,依存句法分析等,使得错误级联情况较为严重;基于最大熵的方法将事件元素识别问题看作分类问题,将所有出现的Entity、Time Expression和Value作为候选元素,选取词法、类别信息、上下文信息、句法结构等特征从不同的角度描述候选元素,并采用最大熵分类器对其进行多元分类。对比实验证明,由于候选元素实例比较充裕,多元分类的方法取得了很好的效果。在ACE2005语料上的实验结果显示,最终的F值达到了64.64%,高于基于SRL的元素识别方法。更多还原

【Abstract】 Event extraction is a very important research point in the area of information extraction. Event extraction can present the event which was describes by natural language through structural form, e.g. who, where, when and what is related to the event. And this technology can be widely applied to many NLP researches, such as summarization, question and answering, information retrieval and so on. This paper makes an intensive study of the two stages of Chinese event extraction: event type recognition and event argument recognition, and then develops a practical event extraction system named HIT-IR EES.In the research of event type recognition, for the reason of small amount of event instances limited by the corpus, the data sparseness problem brought by the small set of training data is the main point of this stage. In this paper, we present a novel method based on automatically extending event triggers to solve this problem in which we first extend event triggers via thesaurus, and extract the candidate events and their candidate types by using extended triggers, and then we adopt a binary classification method to recognize the type of the candidate events. This method solves the data unbalanced problem in training model and the data sparseness problem brought by the small set of training data effectively. Moreover, the precision and the recall of event type recognition are improved. Evaluation on the datasets of ACE2005 shows that, the final F-score achieves 61.24% which outperforms the traditional methods based on machine learning significantly.In the research of event argument recognition, how to recognize the right argument from lots of entity, time expression and value is the main point. In this paper, we present two methods to slove this problem: SRL(Semantic Relation Labeling) based and ME(Maximum Entropy) based. SRL based method sloves the argument recognition problem from application, which tallies with the task of SRL. But for the reason of an over reliance on the bottom techniques, such as SRL and syntax parser, error cascade becomes a big problem; ME based method considers the event argument recognition as a classification problem and considers all the entity, time expression, value which appeares as candidate argument, and then describes these candidate arguments by using lexical, type, context and syntax features from different angles. Finally we adopt a multi classification method to recognize the role of the candidate arguments. The experimental results show that, the multi classification method achieved a better result due to the plenty of candidate argument instances. Evaluation on the datasets of ACE2005 shows that, the final F-score achieves 64.64%.更多还原