Enhancing Knowledge Acquisition Systems with User Generated and Crowdsourced Resources

Abstract

This thesis is on leveraging knowledge acquisition systems with collaborative data and crowdsourcing work from internet. We propose two strategies and apply them for building effective entity linking and question answering (QA) systems. The first strategy is on integrating an information extraction system with online collaborative knowledge bases, such as Wikipedia and Freebase. We construct a Cross-Lingual Entity Linking (CLEL) system to connect Chinese entities, such as people and locations, with corresponding English pages in Wikipedia. The main focus is to break the language barrier between Chinese entities and the English KB, and to resolve the synonymy and polysemy of Chinese entities. To address those problems, we create a cross-lingual taxonomy and a Chinese knowledge base (KB). We investigate two methods of connecting the query representation with the KB representation. Based on our CLEL system participating in TAC KBP 2011 evaluation, we finally propose a simple and effective generative model, which achieved much better performance. The second strategy is on creating annotation for QA systems with the help of crowdsourcing. Crowdsourcing is to distribute a task via internet and recruit a lot of people to complete it simultaneously. Various annotated data are required to train the data-driven statistical machine learning algorithms for underlying components in our QA system. This thesis demonstrates how to convert the annotation task into crowdsourcing micro-tasks, investigate different statistical methods for enhancing the quality of crowdsourced annotation, and finally use enhanced annotation to train learning to rank models for passage ranking algorithms for QA.