Dynamic Chunkwise CNN for Distantly Supervised Relation Extraction

Sentence representation learning is a key component in distantly supervised relation extraction. Text chunks (i.e., a group of n consecutive words) are meaningful units to understand relations. However, existing works suffer from extracting useful structural features from text chunks for relation extraction due to two challenges: (1) Prepositions often occur in text chucks but their semantics are hardly captured. (2) Chunk structures vary in different sentences with different sizes. In this paper, we propose a new model, dynamic chunkwise CNN (DCW-CNN), to tackle these challenges. We develop structural convolution to extract chunk features from sentences, and design a dynamic chunk module to dynamically determine the "most proper"chunk size for sentences of varying structures and contents. We have conducted experiments on two benchmark datasets. The experimental results show that our proposed model improves performance significantly compared with the state-of-the-art methods.