Natural Language Inference Based on the LIC Architecture with DCAE Feature

Natural Language Inference (NLI), which is also known as Recognizing Textual Entailment (RTE), aims to identify the logical relationship between a premise and a hypothesis. In this paper, a DCAE (Directly-Conditional-Attention-Encoding) feature based on Bi-LSTM and a new architecture named LIC (LSTM-Interaction-CNN) is proposed to deal with the NLI task. In the proposed algorithm, Bi-LSTM layers are used to modeling sentences to obtain a DCAE feature, then the DCAE feature is reconstructed into images through an interaction layer to enrich the relevant information and make it possible to be dealt with convolutional layers, finally the CNN layers are applied to extract high-level relevant features and relation patterns and the discriminant result obtained through a MLP (Multi-Layer Perceptron). Advantages of LSTM layers in sequence information processing and CNN layers in feature extraction are fully combined in this proposed algorithm. Experiments show this model achieving state-of-the-art results on the SNLI and Multi-NLI datasets.