Multi-class Named Entity Recognition Via Bootstrapping with Dependency Tree-Based Patterns

Abstract

Named Entity Recognition (NER) has become a well-known problem with many important applications, such as Question Answering, Relation Extraction and Concept Retrieval. NER based on unsupervised learning via bootstrapping is gaining researchers’ interest these days because it does not require manually annotating training data. Meanwhile, dependency tree-based patterns have proved to be effective in Relation Extraction. In this paper, we demonstrate that the use of dependency trees as extraction patterns, together with a bootstrapping framework, can improve the performance of the NER system and suggest a method for efficiently computing these tree patterns. Since unsupervised NER via bootstrapping uses the entities learned from each iteration as seeds for the next iterations, the quality of these seeds greatly affects the entire learning process. We introduce the technique of simultaneous bootstrapping of multiple classes, which can dramatically improve the quality of the seeds obtained at each iteration and hence increase the quality of the final learning results. Our experiments show beneficial results.