低标注资源下的自然语言结构学习

自然语言处理领域近两年出现了一些新的趋势。首先，为了减轻现有方法特别是基于深度神经网络的方法对大量标注数据的依赖，低标注资源下的学习开始受到越来越多的关注。其次，很多研究人员开始重新审视和探索语言结构在各类自然语言处理方法包括神经网络方法中的积极作用。本项目计划结合上述两个趋势，研究低标注资源下的语言结构学习及其对下游任务的作用。本项目将重点研究两类语言结构：句法解析（包括依存解析树和成分解析树）和语义解析（例如语义依存解析图等）。本项目将从不使用标注资源的无监督学习出发，以之为基础再研究使用少量标注资源的少样本学习和基于下游任务反馈的学习等低标注资源学习方式。在以上研究的基础上，本项目计划探索低标注资源下所学得的语言结构对于下游任务的积极作用。本项目的意义在于通过减少对大量标注数据的依赖，使得语言结构可以更方便地被学习和预测，从而促进本领域对语言结构在各类任务中的作用进行重新审视。

There are two new trends in the field of natural language processing. First, many current methods, especially deep neural network methods, rely on large amounts of annotated data; to mitigate this reliance, researchers are increasingly paying attention to learning with few annotated resources. Second, many researchers begin to reinvestigate the utility of linguistic structures in various natural language processing methods including neural methods. Combining these two trends, this project plans to study linguistic structure learning with few annotated resources and its usefulness in downstream tasks. We will focus on two types of linguistic structures: syntactic parses (including dependency parse trees and constituent parse trees) and semantic parses (such as semantic dependency parse graphs). We will first study unsupervised learning that does not require any annotated resource; based on that, we will then study learning methods with few annotated resources, such as few-shot learning (which uses small amounts of annotated data) and learning from feedback of downstream tasks. Based on the above research, we plan to further investigate the utility of linguistic structures in downstream tasks when they are learned with few annotated resources. The ultimate purpose of this project is to ease the learning and prediction of linguistic structures by reducing their reliance on large annotated datasets, so as to facilitate the reinvestigation of the utility of linguistic structures in various tasks.