Residual Connected Enhanced Sequential Inference Model for Natural Language Inference

Understanding the semantic and logical relationships between sentence pair is a difficult problem to be solved in natural language understanding tasks. Although the Enhanced Sequential Inference Model is simple in structure and performs well in SNLI, the limited capacity of the this model limits its further improvement of performance. Inspired by Res-Net, we propose the res-ESIM by introducing the residual connection into the ESIM model to expand the capacity of the ESIM while maintaining properties of simple structure and easy training. We explore the performance of res-ESIM with word embedding and the ability of using the contextual embedding to enhance its performance. In the experiments on SNLI, GloVe is used as word embedding for the convenience of comparing with published models. In the experiments on MultiNLI, the output of BERT-base based on different enhancement methods is used as contextual embedding. The experiment results on SNLI showed that our model achieves competitive performance in all models that haven’t employed additional contextualized word representations and the experiment results on MultiNLI showed that res-ESIM can have more performance improvement than the original ESIM when the information of embedding is enhanced.