Collaborative: Improving Subjectivity Analysis to Achieve High-Precision Information Extraction

University Of Utah, Salt Lake City UT

Investigators

Abstract

This project will use subjectivity analysis to improve the accuracy of information extraction (IE) systems. IE systems are designed to extract facts, but they are prone to false hits from subjective statements such as accusations, allegations, suspicions, and opinions. The first phase of the research will create a subjectivity classifier that uses learning algorithms to identify linguistic features associated with subjective language. The classifier will use several natural language representations, including extraction patterns, N-grams, and noun phrases. The classifier will be embedded in a bootstrapping architecture so that it can learn from unannotated corpora, requiring only a small amount of annotated data to jumpstart the bootstrapping. In the second phase, the classifier will be integrated into an IE system to measure the impact of subjectivity classification on IE performance. Information extracted from objective sentences will be treated as facts, but information extracted from subjective sentences will be labeled as uncertain or discarded. This research will produce a better understanding of how subjective language is expressed and the role that context plays in recognizing subjectivity. The potential impact of the research is to produce more accurate subjectivity classifiers and to demonstrate that subjectivity analysis can improve the performance of IE systems.