Syntax–Aware graph convolutional network for the recognition of chinese implicit inter-sentence relations

In the literature, most previous studies on English implicit inter-sentence relation recognition only focused on semantic interactions, which could not exploit the syntactic interactive information in Chinese due to its complicated syntactic structure characteristics. In this paper, we propose a novel and effective model DSGCN-RoBERTa to learn the interaction features implied in sentences from both syntactic and semantic perspectives. To generate a rich contextual sentence embedding, we exploit RoBERTa, a large-scale pre-trained language model based on the transformer unit. DSGCN-RoBERTa consists of two key modules, the syntactic interaction and the semantic interaction modules. Specifically, the syntactic interaction module helps capture the depth-level structure information, including non-consecutive words and their relations, while the semantic interaction module enables the model to understand the context from the whole sentence to the local words. Furthermore, on top of such multi-perspective feature representations, we design a strength-dependent matching strategy that is able to adaptively capture the strong relevant interactive information in a fine-grained level. Extensive experiments demonstrate that the proposed method achieved state-of-the-art results on benchmarks Chinese compound sentence corpus CCCS and Chinese discourse corpus CDTB datasets. We also achieve comparable performance on the English corpus PDTB that demonstrates the superiority of our method.