CAREER: Cross-Document Cross-Lingual Event Extraction and Tracking

本研究项目的目标是将信息抽取的范式推进到“空位填充”之外，通过利用动态背景知识和跨文档的跨语言事件排序和跟踪来获得更准确、突出、完整、简洁和连贯的抽取结果。该方法包括跨文档推理、未知隐式事件时间预测和推理、全局上下文的跨文档实体共指消解、质心实体检测、事件属性提取和基于图的聚类算法的冗余和冲突检测、新事件的自动聚类和主动学习、基于提取结果的摘要生成、具有可比性的语料库名称翻译和跨语言协同训练。该项目包括博士生和本科生，非计算机科学本科生参与效用评估和语料库注释，并通过教程、定期研究研讨会和广泛的暑期研讨会吸引小学生和高中生。本项目的成果也将通过从中小学使用的科学文献和学习材料中提取和跟踪相关知识而在E-Science和E-Learning中受益。项目成果，包括开源软件、任务定义指南、注释语料库、评分指标将通过项目网站(http://nlp.cs.rpi.edu/blendeet.html).

The goal of this research project is advance the Information Extraction (IE) paradigm beyond "slot filling", and achieve more accurate, salient, complete, concise and coherent extraction results by exploiting dynamic background knowledge and cross-document cross-lingual event ranking and tracking. The approach consists of cross-document inference, unknown implicit event time prediction and reasoning, cross-document entity coreference resolution with global contexts, centroid entity detection, event attribute extraction and graph-based clustering algorithms for redundancy and contradiction detection, automatic new event clustering and active learning, abstractive summary generation based on extraction results, name translations with comparable corpora and cross-lingual co-training.The experimental research is integrated with educational activities, including project-related curriculum development. The project involves PhD students as well as undergraduate students, engages non-Computer Science undergraduate students in utility evaluation and corpus annotation, and attracts elementary school and high school students by tutorials, regular research seminars and an extensive summer workshop. The results of this project will also have a benefit in E-Science and E-Learning by extracting and tracking the related knowledge from scientific literature and learning materials used in elementary schools and high schools.Project results, including open source software, task definition guidelines, annotated corpora, scoring metrics will be disseminated via project Web site (http://nlp.cs.rpi.edu/blendeet.html).