Neural Event Extraction from Scientific Literature

本文的主要目的是研究能够从文本中提取多个结构的事件提取模型。虽然已经提出了许多用于事件提取的神经模型，但大多数都集中在平面事件和平面实体上。事件提取的最新方法是管道系统，它独立地按顺序处理小任务，以解决更复杂的任务。然而，与以联合方式处理子任务相比，顺序处理可能导致整体性能不佳。这可以归因于这样一个事实，即由管道的任何早期子任务引起的错误都会传播到系列中的其余子任务。此外，基于管道的方法无法捕获不同子任务之间的相互依赖关系。因此，在管道中单独处理的元素（例如，命名实体和事件）不能用于相互消歧。只有少数事件提取方法使用了嵌入，并且只使用单词的词汇和句法特征。我们将研究用于事件提取的新型神经网络结构化预测模型，建立在上下文嵌入之上，同时检测实体/触发器、角色层、触发器-参数对和事件层。我们的目标是最大限度地提高精度和高覆盖率。通过将深度信息提取分解为独立的组件，我们可以将每个组件视为一个独立的系统，以支持组件之间的轻松连接。这些组件的模型将支持联合学习，因此各个模型必须共享表示。为了达到高精度，我们将在事件提取管道的不同组件之间部署联合训练，并将多任务学习与其他任务（例如POS标记，语言模型）结合起来。我们还将处理复杂的数据表示，如嵌套和不相交的实体、触发器和事件，并考虑词之间的语义关系结构以提高性能。为了实现高覆盖率，我们将研究语言单位（如字符、子词、短语和多词表达）的联合表示，这将使从外部人工管理资源（如知识库、本体、词典和数据库）中更灵活和联合学习成为可能。

The main objective of the thesis is to investigate event extraction models able to extract multiple structures from text. Although many neural models have been proposed for event extraction, most have focused on flat events and flat entities. State-of-the art approaches to event extraction are pipeline systems, which independently addresses small tasks in sequence, to solve a more complex task. However, sequential processing may result in suboptimal overall performance when compared to approaching the subtasks in a joint manner. This can be attributed to the fact that errors resulting from any of a pipeline's early subtasks are propagated to the rest of the subtasks in the series. Also, pipeline-based approaches fail to capture inter-dependencies amongst the different subtasks. As a result, elements that are separately dealt with in a pipeline (e.g., named entities and events) cannot be used for mutual disambiguation. Only a handful of event extraction approaches have employed embeddings, and using only lexical and syntactic features of words.We will investigate novel neural network structured prediction models for event extraction, building on top of the contextual embeddings and simultaneously detecting entity/triggers, role layers, trigger-argument pairs and event layers. Our aim is to maximise both precision and high-coverage. By decomposing deep information extraction into separate components, we can consider each one as an individual system to support easy connectivity between components. The models of these components will support joint learning, thus the individual models have to share representations. To achieve high precision we will deploy joint training between different components of the event extraction pipeline and we will incorporate multi-task learning with other tasks, e.g. POS tagging, language models. We will also deal with complex data representations such as nested and disjoint entities, triggers and events and take into account semantic relational structure between words to improve performance. To achieve high coverage, we will investigate joint representation of linguistic units such as characters, sub-words, phrases and multi-word expressions which will enable more flexible and joint learning from external manually curated resources such as knowledge bases, ontologies, lexica and databases.