Analysis of Deep Neural Networks for Textual Entailment Recognition
Purnendu Kumar Rath, Rohini Basak · 2020 Fourth International Conference on Inventive Systems and Control (ICISC) · 2020
Textual entailment methods recognize, pairs of natural language expressions such that a human who reads (and trusts) the first element of a pair would most likely infer that the other element is also true. A lot of research works has already been accomplished in this area. In this paper we analyze the impact of various deep neural network architectures on the performance of recognizing textual entailment pairs. We start with a very simplistic neural network architecture and gradually add elements to the architecture that can possibly improve its performance, thereby analyzing its impact on the performance parameters. We measure the performance based on accuracy, recall, precision, and F1 score of train as well as test data, obtained from SNLI data-set. Here we start by tokenizing and converting the text and the hypothesis into corresponding sequence of word-embedding vectors and comparing these two sequence by passing them through various neural network layers and using a dense layer with softmax activation as the final layer. We construct every new architecture by adding new layers and/or inputs to the previous best architecture. In this paper we try six architectures and analyze their performance in terms of the performance parameters. We analyze the impact of various layers like LSTM, concatenation, cosine similarity function, normalization, dropouts and attention.