Dynamic Knowledge Integration for Natural Language Inference

Natural language inference (NLI) aims to determine the entailment relationship between the premise and the hypothesis. It is a fundamental but difficult problem, since there may exists serious semantic and logistic gap between the premise and the hypothesis. Despite using strong pre-trained language model (PLM), previous work performs poorly on complicated reasoning for knowledge-sensitive cases ignoring the integration of external knowledge. We propose a dynamic knowledge integration strategy for NLI, where knowledge from multiple knowledge graphs (KGs) can be dynamically integrated. For each KG, it transforms input tokens into a graph according to the connectivity of the related entities. All the graphs are encoded by a group of parallel graph neural networks (GNNs), and after each layer the intermediate results are integrated dynamically by being conditioned on the input text. This strategy also facilitates the incorporation of PLM, simply by treating the input tokens as a fully connected graph and adapting the PLM outputs as the node embeddings. Experiments on SNLI, MNLI and SciTail show that, the dynamic integration of knowledge from WordNet and ConceptNet achieves significant improvements over the strongest baseline built upon RoBERTa.