A fine-grained causality extraction model incorporating relative location coding

Abstract

Popular methods of causality extraction work well for simple and explicit single causal relations, but it remains challenging to extract causal relations from the complex sentences of natural texts due to ambiguity concerning the locations of the causal subject and object as well as the complexity of the relevant dependencies. To solve these problems, this paper proposes a five-tuple annotation scheme that defines a scoring function to iteratively parse out the causal pairs from multiple entity pairs in a sentence, and to thus transform the task of extracting causal relations into that of automatic annotation. First, this study uses this scheme to propose a multi-headed, self-attentive mechanism that incorporates encoded information on relative position to increase the capability of the model to perceive causal features. Second, the authors combine information from a dependency tree while assigning the appropriate weights, and finally use a bidirectional GCN network to parse the weights of features of the tree from multiple perspectives and splice the dependency-related features. This joint model of extraction improves the bounds of cause–effect pairs of entities while considering the dependency relationships between them, which renders the extracted, fine-grained causal terms more accurate. Experiments on the SemEval 2010 task 8 and the ADE datasets show that our approach significantly outperforms prevalent methods in terms of the accuracy of solving complex causal extraction compared with state-of-the-art approaches to modeling.