短剧用户观感评价情感分析

Sentiment Analysis of the Users of Micro-Dramas

摘要: 短剧随着时代发展逐渐崛起，成为当今国内外新潮的娱乐载体。本文爬取腾讯短剧品牌十分剧场的短剧用户评价，对该不平衡样本数据进行情感分析，比较多种模型与模型组合的效率与效果。1) 使用Word2vec的连续词袋模型将预处理后的文本转为词向量，构建LSTM/BILSTM模型，两者无效果差别，LSTM所用时间最短；2) 构建TextCNN + LSTM/BILSTM模型，使用TextCNN获取向量特征，通过LSTM/BILSTM学习情感规律，稀少数据的F1-Score提升约10%；3) 构建TextCNN + LSTM + Muti_Head_Attention模型，添加多头注意力机制把握字与字之间的多重联系，耗时增加一倍，稀少数据的F1-Score上限再次提升1%；4) 使用随机删除增强数据会以降低20%的精准率的代价提高10%的召回率；5) 在第3点的基础上在卷积层中添加残差连接，稀少数据的F1-Score上限提高2%；6) 使用Bert/Roberta的分词器和模型取代Word2vec与传统RNN，得到的结果对比第5点，提升约为9%/12%，泛化性更强，时间和硬件成本大幅提升，但添加TextCNN、LSTM与多头注意力后，效果反而出现下降。

Abstract: As Micro-Dramas grow in popularity worldwide, this article evaluates user reviews from Tencent’s “Shifen Theater”, analyzing imbalanced data sentiment and comparing various models and combinations. 1) Word2Vec’s bag-of-words model turns preprocessed text into vectors, building LSTM/BiLSTM models—both perform poorly, with LSTM being the fastest; 2) The TextCNN + LSTM/ BiLSTM model uses TextCNN for vector features and LSTM/BiLSTM for sentiment learning, boosting the F1-Score for rare data by about 10%; 3) Adding Multi-Head Attention to TextCNN + LSTM/BiLSTM captures intricate character relationships, doubling the runtime and increasing the F1-Score by 1%; 4) Random deletion enhances data but sacrifices 20% precision for 10% better recall; 5) Add residual connections to the convolution layers in model 3, improving the F1-Score by 2% on sparse data; 6) Replacing Word2Vec and traditional RNNs with Bert/Roberta improves results by 11%/14% over the third model, offers better generalizability, but increases time and cost significantly. However, incorporating TextCNN, LSTM, and Multi-Head Attention can decrease performance.