Making the Relation Matters: Relation of Relation Learning Network for Sentence Semantic Matching

Sentence semantic matching is one of the fundamental tasks in natural language processing, which requires an agent to determine relation among input sentences. Recently, deep neural networks have achieved impressive performance this area, especially BERT. Despite effectiveness these models, most them treat output labels as meaningless one-hot vectors, underestimating information and guidance relations that reveal, for with a small number labels. To address problem, we propose Relation Learning Network (R2-Net) sentence matching. Specifically, first employ BERT encode sentences from global perspective. Then CNN-based encoder designed capture keywords phrase local fully leverage better extraction, introduce self-supervised classification task guiding R2-Net consider more about Meanwhile, triplet loss employed distinguish intra-class inter-class finer granularity. Empirical experiments on two demonstrate superiority our proposed model. As byproduct, released codes facilitate other researches.