RI: Small: Semantics-Based, Weakly-Supervised Coreference Resolution

本研究试图通过实现两个目标来推动共指消解研究的前沿。首先，它通过考察一类难以解决的回指来应对Winograd图式挑战，这些回指的解决需要常识知识，涉及到参与者在事件中所扮演的角色及其因果关系。它对这一问题采取了深度文本理解的方法。具体地说，它涉足共指研究的未探索领域，包括脚本类知识和情感分析的使用，以及对语篇连接词作用的审查。其次，它使得能够获得比目前可能的数量多得多的自然语言和领域的共指解析器。在大量语言和领域中部署共指技术的主要障碍之一是，在一种语言和领域中对数据进行共指注释的相关成本很高。它研究了两种经济高效的数据标注方法，一种涉及基于翻译的投影，另一种涉及自举。由于共指是许多传统和新兴文本处理应用程序的一种使能技术，该项目具有改进这些应用程序的潜力。通过建立一个多语种、多领域的指代解析器和提供在本调查过程中产生的注释数据，该项目可能会促进更多的研究人员对研究不足的语言和领域进行研究。同样重要的是，在解析器中使用常识知识可以模拟人类共指解析过程，使人工智能研究人员离构建能够真正理解自然语言的智能代理又近了一步。

This research seeks to push the frontiers of coreference resolution research via achieving two objectives. First, it addresses the Winograd Schema Challenge by examining a class of difficult-to-resolve anaphors whose resolution requires commonsense knowledge involving the roles played by the participants in an event and their causal relationships. It adopts a deep text-understanding approach to this problem. Specifically, it ventures into unexplored areas of coreference research, including the use of script-like knowledge and sentiment analysis, as well as an examination of the role of discourse connectives. Second, it enables the acquisition of coreference resolvers for a substantially larger number of natural languages and domains than is currently possible. One of the major obstacles to the deployment of coreference technologies across a large number of languages and domains is the high cost associated with coreference-annotating data in a language and domain. It investigates two cost-effective approaches to data annotation, one involving translation-based projection and the other bootstrapping.As coreference is an enabling technology for many traditional and emerging text-processing applications, the project has the potential to improve these applications. Through the construction of a multi-lingual, multi-domain coreference resolver and the availability of annotated data produced in the course of this investigation, the project may stimulate research in under-studied languages and domains by a broader community of researchers. Equally importantly, the use of commonsense knowledge in the resolver mimics the human coreference resolution process, bringing artificial intelligence researchers one step closer to building an intelligent agent that can truly understand natural language.