Text Mining and Analysis Based on XLNet-BiLSTM Model

Abstract

In the light of the BERT model relies on masking part of the word to achieve bi-directional prediction of text, but does not take into account the possible association between masked and unmasked words, resulting in loss of information and long-term dependency and inadequate context-deep semantic mining in RNN, an emotion classification model based on XLNet BiLSTM is used. This paper first uses the XLNet pre-training language model to generate context-dependent word vectors to represent text information in a distributed way; Then the word vector is input into the BiLSTM network to analyze and calculate the deep semantics of the text; Finally, softmax function is used to output emotional polarity classification. Train and test the model by mining the text about the Lantern Festival on some social and media platforms. The experimental results show that the accuracy rate of the model is 0.958, and the loss rate is 0.182.