A multi-level neural network for implicit causality detection in web texts

Abstract

Mining causality from text is a complex and crucial natural language understanding task corresponding to human cognition. Existing studies on this subject can be divided into two categories: feature engineering-based and neural model-based methods. In this paper, we find that the former has incomplete coverage and intrinsic errors but provides prior knowledge, whereas the latter leverages context information but has insufficient causal inference. To address the limitations, we propose a novel causality detection model named MCDN, which explicitly models the causal reasoning process, and exploits the advantages of both methods. Specifically, we adopt multi-head self-attention to acquire semantic features at the word level and develop the SCRN to infer causality at the segment level. To the best of our knowledge, this is the first time the Relation Network is applied with regard to the causality tasks. The experimental results demonstrate that: i) the proposed method outperforms the strong baselines on causality detection; ii) further analysis manifests the effectiveness and robustness of MCDN.