低资源依存句法模型构建与可解释性分析

作为一类简单有效的文本结构化方法，依存句法分析对信息抽取，语义理解等任务都有重要作用，是自然语言处理基础任务的核心研究方向之一。随着文本分析技术在智能系统中广泛应用，如何扩展依存句法分析应用场景，使其更有效的处理多种语言和多种领域的数据，特别是标注数据稀少的语言和领域，成为了当前依存句法分析的关键问题。..本项课题中，我们将关注面向低资源的依存句法分析。研究统一低资源与高资源的依存分析框架，跨语跨领域场景下依存分析模型与算法。同时，我们也将探究依存分析模型可解释性，研究语言学知识指导下的依存分析错误诊断与异常检测，提升低资源依存分析的稳定性与鲁棒性。具体研究目标包括：1面向依存分析的词表示；2融合异构语料的依存分析编码器；3基于联合学习的依存分析解码器； 4低资源依存分析的可解释性。..预期的研究成果包括：发表国内外学术期刊或会议论文15-20篇，申请专利3-4项，开源低资源依存分析工具。

Being of a simple and effective method for detecting structural information in texts, dependency parsing plays an important role in many information extraction and natural language understanding tasks. Traditionally, dependency parsers are trained and tested on Treebanks with documents from news and official documents in major languages(e.g., English). However, it is always the case that for other languages and text domains, enough annotations of dependency trees are hard to obtain. How to parsing with low resources is therefore crucial to enlarge the application of dependency grammar and attracts recent research attentions...In this proposal, we focus on building a unified dependency parsing model for both high and low resources, and we also propose to investigate the interpretability of the current parsing model which helps us understanding the statistical model better. We hope that by exploring cutting edge cross-lingual cross-domain algorithms and model verification methods, we are able to improve the robustness of dependency parsing, understand the behaviors of them better and benefit various NLP applications.