Sentiment Analysis of Barrage Text Based on ALBERT-ATT-BiLSTM Model
Yifan Wu, Jianjun He · 2021
For the sentiment analysis of barrage texts, traditional methods cannot distinguish the different meanings of the same word in a sentence in different contexts when performing feature extraction, which cannot take into account the local feature information in the text and the contextual semantic association during the training process. As a result, its classification accuracy is relatively low. To this end, this paper proposes a sentiment classification model for barrage text that combines the ALBERT pre-training language model and BiLSTM with the attention mechanism. First, use the ALBERT pre-training language model to obtain the dynamic feature representation of the barrage text, then use the BiLSTM network to extract the text context relationship features, dynamically adjust the feature weights through the attention mechanism, and then use the Softmax classifier to obtain the sentiment category of the barrage text. Experiments on the barrage text data set obtained by the crawler show that the ALBERT-ATT-BiLSTM model compares W2V-Att-CNN and W2V-Att-LSTM models with attention mechanism, the precision value has increased by 5.34% and 4.32%, the recall value has increased by 5.23% and 4.27%, and the F1 has increased by 5.27% and 4.25%. Compared with the MC-CNN-GRU model, the precision value of ALBERT -ATT -BiLSTM model is also improved by 3.61%, which is more accurate on the data set than some traditional models.