基于LSTM和多特征组合的电影评论专业程度分类

Movie Review Professionalism Classification Using LSTM and Features Fusion

社交网络上的电影评论通常既有专业评论家写的专业评论,也有普通观众写的非专业评论,区分网络电影评论是否为专业评论对于电影质量评估有着重要的价值。由于电影评论属于短文本,用词不规范,特征稀疏,因此传统的文本特征选择方法以及传统的分类模型并不能完全适用于电影评论专业程度的分类。为此,文中主要研究基于神经网络模型电影评论的专业程度分类,即判断其是专业评论还是非专业评论。首先通过基于神经网络的LSTM模型学习不同特征的表示,包括基于词的表示、基于词性的表示,以及基于依存关系的表示,然后通过融合不同特征表示来学习和捕捉有效的文本特征,从而帮助评论专业程度分类。该方法在美国著名的影评网站烂番茄网(Rotten Tomatoes)数据集上进行实验,实验结果表明,在融合了词性和依存关系特征的模型的分类正确率达到了88.30%,比仅使用词特征的基准模型提高了3.66%。这说明在模型中引入词性特征、依存关系特征能够有效提升评论专业程度分类的效果。

Movie Reviews on social networks usually include professional reviews written by professional critics,as well as non-professional reviews written by ordinary audience,and it is of great value to distinguish whether online film reviews are professional reviews for film quality evaluation.Due to the fact that film review is a short text book with irregular words and sparse features,the traditional text feature selection method and traditional classification model cannot fully apply to the classification of film review’s professional level.Therefore,the paper mainly studied movie review professionalism classification based on neural network model,that is judging whether it is professional review or non-professional review.The representation of different features is learned through neural network-based LSTM model,including word-based representation,part-of-speech representation,and representation based on dependencies,and valid text features are learned and captured by fusing different feature representations to help review professionalism classification.The method was experimented on the Rotten Tomatoes dataset of the famous American film review website.The experimental results show that the classification accuracy rate of the model combining part-of-speech and dependency is 88.30%,which is 3.66%higher than the benchmark model only using word features.This shows that the method of introducing part-of-speech features and dependency features into the model can effectively improve the effectiveness of professional classification of reviews.