毕业生知识库

在信息时代，互联网已经成为人们获取信息的主要渠道，网络上的信息规模也呈爆炸式增长。信息抽取技术旨在从非结构化的自然语言文本中抽取出结构化的事实描述，从海量文本数据中挖掘关键信息。事件抽取作为信息抽取领域中的关键环节，目标是从自然语言文本中识别并抽取出结构化事件信息。近年来，在相关国际评测的推动下，事件抽取研究取得了较大进展。但是，目前大多数事件抽取方法都是面向句子层面的，即分析单个语句的文本内容，并从中抽取事件知识。相对而言，面向篇章级的事件抽取方法研究较少，方法技术十分有限。在实际应用场景中，事件在文本中的表达通常是篇章级别的。相对于句子级别的事件表达，篇章级别的事件表达更加复杂和多样，给篇章级事件抽取技术带来了极大的挑战。

本文重点关注篇章级事件抽取方法，围绕篇章级事件抽取过程中普遍存在的单事件元素跨句分散问题、多事件交错描述问题和主事件多视角表达问题展开研究。
本文的主要研究内容和创新点如下：

1.针对篇章内单事件元素跨句分散问题，本文提出了一种基于层级编码的跨句事件元素抽取方法。该方法利用层级编码器有效地建模了篇章中局部（句内）和全局（句间）的信息交互，并获得篇章级语义表示，从而实现了跨句事件元素的抽取。具体地，针对事件元素与触发词存在跨句联系的场景，设计了一个基于事件感知和层级编码的抽取模型，隐式地捕获事件文本信息的交互。该模型采用层级编码器建模篇章上下文和事件文本之间的局部和非局部依赖关系，并引入事件内容预测的辅助任务来帮助元素抽取模型隐式地感知事件相关的信息。同时，针对事件元素之间存在跨句联系的场景，设计了一个基于多轮多粒度阅读框架的抽取模型，显式地建模事件元素之间的关系。该模型基于机器阅读理解的范式，为事件元素抽取提供事件语义角色的文本信息，并引入多轮问答的形式捕获事件元素之间的长距离依赖关系。在标准篇章级事件元素抽取数据集上的实验结果表明，相比以往的抽取方法，上述方法可以显著提升事件元素抽取的性能，尤其在跨句元素抽取这种具有挑战性的场景中表现更为突出。

2.针对篇章内多事件交错描述问题，本文提出了一种基于并行预测的多事件协同抽取方法。该方法基于编码器-解码器的生成式框架，有效地融入篇章级信息，建模多事件之间的信息交互，实现多个结构化事件的并行预测。具体地，在模型层面，该方法首先引入文档级编码器进行篇章级信息融合，捕获篇章内容感知的上下文表示。其次，设计了一个基于非自回归的多粒度事件解码器，该解码器不仅能够建模多事件间的信息交互，还支持多个结构化事件的并行预测。最后，在模型优化层面，引入了基于二部图匹配算法的损失函数，实现了端到端生成式模型的全局优化。在标准篇章级事件抽取数据集上的实验结果表明，相对于以往的抽取方法，该方法可以显著提升篇章级事件抽取的性能，特别是在具有挑战性的多事件抽取场景中性能提升尤为明显。

3.针对篇章内主事件的多视角表达问题，本文提出了一种基于篇章结构建模的主事件多视角分析方法。该方法通过多视角语篇图的构建，有效地建模篇章结构，实现面向事件的语篇内容分析。具体地，该方法首先构建以主事件为导向的多视角语篇图来建模篇章结构。多视角图由注入先验知识的语篇关系图、共指关系图和基于潜在图结构学习的动态图三部分构成，用于捕获语篇单元之间的语义关系。其次，利用图卷积神经网络对多视角图结构中的语义内容进行表示学习。
最后，进一步引入动态路由网络来分配从多视角特征表示到内容预测的贡献权重。在面向事件的内容结构分析数据集上的实验结果表明，相比以往的方法，本文所提方法可以显著提升篇章事件分析的性能。

本文的工作、方法和结论对于进一步探索和建立更加高效的篇章级事件抽取系统具有重要的参考价值和指导意义。

With the rapid development of information technology, the amount of information on the internet has increased exponentially. Information extraction (IE) technology aims to extract structured fact descriptions from unstructured natural language texts and mine valuable information from massive textual data. As a key component of IE, event extraction (EE) aims to identify and extract structured event information from texts. In recent years, driven by international evaluations, EE research has made great progress. However, despite many efforts for EE, most of the existing EE methods focused on the sentence-level, which analyzes event information expressed in one sentence and extracts event knowledge from it. In real-world scenarios, events are typically expressed at the document-level, which is more complex, diverse, and varied. These characteristics also pose significant challenges for document-level event extraction technology.

This thesis is centered on document-level EE, which aims to address the common issues faced during the document-level EE process. 
 These issues include scattered event arguments across sentences, interlaced descriptions of multiple events, and multi-perspective expressions of the main event. The main research contents and innovative points of this paper are as follows:

1. To addresses the problem of scattered event arguments across sentences in a document, this thesis proposes a hierarchical encoding-based argument extraction method. The hierarchical encoder helps to deeply understand the document content, allowing for the modeling of both local (intra-sentence) and global (inter-sentence) information interactions within the document.  To address scenarios where arguments distribute across sentences with trigger words, we propose an event-aware hierarchical encoder for multi-sentence argument linking, which can capture the interactions in a textual event implicitly.  Specifically, a hierarchical encoder is introduced to capture local and global interactions in a textual event, and an auxiliary task is introduced to predict the event-relevant context, which benefits the argument-linking model by making it aware of the event-relevant context. Furthermore, to address scenarios where arguments distribute across sentences, we introduce a multi-turn and multi-granularity reader to capture the dependencies between arguments explicitly. This model is based on the paradigm of machine reading comprehension (MRC), providing textual information on semantic roles for argument extraction and introducing a multi-round question-and-answer format to explicitly capture long-distance dependencies between event arguments. Experimental results on standard document-level event argument extraction datasets demonstrate that the proposed methods significantly outperform current state-of-the-art methods, especially in challenging cross-sentence argument extraction scenarios.

2. To address the problem of interleaved descriptions of multiple events in a document, this thesis proposes a parallel prediction-based method for the collaborative extraction of multiple events.  The method utilizes an encoder-decoder generative framework for parallel predictions of multiple structured events.  Specifically, we first introduce a document-level encoder to integrate the document-aware representations. Then, a multi-granularity non-autoregressive decoder is used to model the interaction between multiple events and generate structured events in parallel. Finally, to train the end-to-end generative model, a matching loss function is proposed, which can bootstrap a global optimization. Empirical results on the standard document-level EE dataset demonstrate that our approach significantly outperforms current state-of-the-art methods, particularly in challenging scenarios of extracting multiple events.

3. To addresses the problem of multi-perspective expression of the main event in a document, this thesis proposes a structure modeling-based method for multi-perspective analysis around main events. Based on the theory that discourse units are typically organized as a global structure with respect to the central events in a news article, this method constructs a multi-perspective discourse graph to model discourse structures. Specifically, we propose multi-perspective graphs to utilize the global structure of a document, which is designed to inject prior knowledge and induce latent structural variables.  Over these rich structural multi-perspective graphs, Graph Convolutional Networks (GCNs) are introduced for representation learning. Additionally, since these graphs model the document-level structure from various perspectives, we further introduce a routing stage to dynamically establish weights between multi-perspective representations and output content types. Experimental results on the discourse profiling dataset demonstrate that this method can significantly improve the performance of event-oriented discourse analysis. The methods and conclusions of this thesis shed light on the exploration and establishment of a more efficient document-level event extraction system.