基于ALBERT-CRNN的弹幕文本情感分析

Barrage Text Sentiment Analysis Based on ALBERT-CRNN

提出一种结合ALBERT预训练语言模型与卷积循环神经网络(convolutional recurrent neural network,CRNN)的弹幕文本情感分析模型ALBERT-CRNN。首先使用ALBERT预训练语言模型获取弹幕文本的动态特征表示,使得句子中同一个词在不同上下文语境中具有不同的词向量表达;然后利用CRNN对特征进行训练,充分考虑了文本中的局部特征信息和上下文语义关联;最后通过Softmax函数得出弹幕文本的情感极性。在哔哩哔哩、爱奇艺和腾讯视频三个视频平台的弹幕文本数据集上进行实验,结果表明,ALBERT-CRNN模型在三个数据集上的准确率分别达到94.3%、93.5%和94.8%,相比一些传统模型具有更好的效果。

The barrage text sentiment analysis model ALBERT-CRNN based on ALBERT pre-training language model and convolutional recurrent neural network(CRNN)was proposed.Firstly,the ALBERT pre-training language model was used to obtain the dynamic feature representations of barrage texts,so that the same word had different word vector expressions in different contexts.Then,these feature vectors were trained by CRNN,which made the local features and context semantic correlation to be fully considered.Finally,the sentiment polarity of barrage texts was obtained by the Softmax function.Experiments were carried out on the barrage text datasets of Bilibili,iQiYi and Tencent video platforms.The experimental results showed that the accuracy of ALBERT-CRNN on the above three datasets reached 94.3%,93.5%and 94.8%respectively,which were better than some traditional models.