Fusing Distinguish Degree Neural Networks for Relational Classification

Abstract

Relational classification is an important natural language processing (NLP) task. Conventional relational classification models predominantly rely on the shortest dependency paths (SDPs) obtained by NLP tools to obtain keywords information or introduce lexical-level extra information. Bidirectional long short-term memory neural network (Bi-LSTM) is a generally accepted and effective method for obtaining complete sentence information. This model uses the forward neural network and backward neural network encoding cascade on step as output results. However, for relational classification task that considers the direction of entities, direction is critical for good experimental results. In fact, the traditional Bi-LSTM cannot clearly distinguish between two relationships with only different entities directions, such as Cause-Effect (e1,e2) and Cause-Effect (e2,e1). Thus, we propose a model called Di-LCNN, which introduces a concept of distinguish degree to distinguish between two types of relationships that differ only in the direction of the entities. In the proposed model, the distinguish degree is defined as the vectorized representation of the relationship between two entities. After the distinguish degree is obtained, it is integrated into the model to enhance the ability of the model to distinguish the categories of different entity directions in the same relationship. The results of experiments conducted on the SemEval-2010 Task 8 dataset, which is designed for nominal relational classification tasks, indicate that our model is an improvement over the state-of-the-art on this dataset, achieving an F1 score of 85.6% without using any manual features or NLP tools. In addition, our experimental results show that (1) distinguish degree is a very important information for relational classification tasks; (2) distinguish degree can greatly improve the results of the basic model.