Extracting Relations from the Web via Weakly Supervised Learning

In the time of big data, information extraction at a large scale has been an important topic discussed in natural language processing and information retrieval. Specifically, weak supervision, as a novel framework that need not any human involvement and can be easily adapted to new domains, is receiving increasing attentions. The current study of weak supervision is intended primarily for English, with conventional features such as segments of words based lexical features and dependency based syntactic features. However, this type of lexical features often suffer from the data sparsity problem, while syntactic features strongly rely on the availability of syntactic analysis tools. This paper proposes to make use of n-gram features which can relieve to some extent the data sparsity problem brought by lexical features. It is also observed that the n-gram features are important for multilingual relation extraction, especially, they can make up for the syntactic features in those languages where syntactic analysis tools are not reliable. In order to deal with the quality issue of training data used in weakly supervised learning models, a bootstrapping approach, co-training, is introduced into the framework to improve this extraction paradigm. We study the strategies used to combine the outputs from different training views. The experimental results on both English and Chinese datasets show that the proposed approach can effectively improve the performance of weak supervision in both languages, and has the potential to work well in a multilingual scenario with more languages.