融合关系类型的图卷积网络关系抽取模型

Relation Extraction Model Based on Graph Convolutional Network with Relations Types

摘要: 关系抽取是信息抽取和检索中的一项重要任务，它旨在从运行的文本中提取给定实体之间的关系。以前的研究表明，要想在这个任务中取得良好的表现，需要对上下文信息进行良好的建模，其中输入句子的依存树可以成为不同类型上下文信息中的一个有益来源。然而，这些研究大多集中在词与词之间的依赖关系上，对依赖类型的利用关注有限。现有的研究大多存在依存树中存在噪声的问题，尤其是在自动生成依存树时，大量利用依赖信息可能会给关系分类带来混乱，因此对依存树进行必要的修剪非常重要。此外，他们在建模时经常平等地对待不同的依赖连接，因此在自动生成的依存树中会受到干扰(不准确的依赖解析)。该论文提出了一种用于关系提取的注意力机制的图卷积神经网络方法，该方法将基于图卷积网络的注意机制应用于依赖解析器获得的依存树中不同上下文单词，以区分不同单词依赖的重要性。又加入一个新定义的模块，将其命名为Key-Value Slot (简称KV Slot)。对于实体中的每个单词，KV Slot模块将所有关联的单词及其之间的依赖性进行映射，然后根据对关系提取的贡献为其分配一个权重。该方法不仅利用了单词之间的依赖连接和类型，而且还将可靠的依赖信息与嘈杂的信息区分开来，在此基础上并对它们进行适当的建模。在SemEval2010-Task8和KBP37数据集上的实验证明了我们的方法的有效性，模型在性能上有了较大提升。

Abstract: Relation extraction is an important task in information extraction and retrieval, which aims to ex-tract relations between given entities from running text. Previous studies have shown that good performance in this task requires good modeling of contextual information, where the dependency tree of the input sentence can be a beneficial source among different types of contextual infor-mation. However, most of these studies focus on word-to-word dependencies and pay limited atten-tion to the exploitation of dependency types. Most of the existing studies have the problem of noise in the dependency tree, especially when the dependency tree is generated automatically, a large amount of dependency information may bring confusion to the relation classification, so it is very important to do necessary pruning of the dependency tree. Moreover, they often treat different de-pendency connections equally when modeling, and thus suffer from interference (inaccurate de-pendency parsing) in the automatically generated dependency tree. This paper proposes a graph convolutional neural network method for the attention mechanism of relation extraction, which ap-plies the attention mechanism based on graph convolutional network to different context words in the dependency tree obtained by the dependency parser to distinguish the importance of different word dependencies. This paper adds a new module called the Key-Value Slot (KV Slot for short). For each word in an entity, the KV Slot module maps all associated words and dependencies between them, and then assigns it a weight based on its contribution to relation extraction. The proposed method not only exploits the dependency connections and types between words, but also distin-guishes reliable dependency information from noisy information, builds on it and models them ap-propriately. Experiments on Semeval2010-task8 and KBP37 datasets prove the effectiveness of our method, and the performance of the model has been greatly improved.