Embedding Logic Rules Into Recurrent Neural Networks

Incorporating prior knowledge into recurrent neural network (RNN) is of great importance for many natural language processing tasks. However, most of the prior knowledge is in the form of structured knowledge and is difficult to be exploited in the existing RNN framework. By extracting the logic rules from the structured knowledge and embedding the extracted logic rule into the RNN, this paper proposes an effective framework to incorporate the prior information in the RNN models. First, we demonstrate that commonly used prior knowledge could be decomposed into a set of logic rules, including the knowledge graph, social graph, and syntactic dependence. Second, we present a technique to embed a set of logic rules into the RNN by the way of feedback masks. Finally, we apply the proposed approach to the sentiment classification and named entity recognition task. The extensive experimental results verify the effectiveness of the embedding approach. The encouraging results suggest that the proposed approach has the potential for applications in other NLP tasks.