摘要： 产品评论的快速增长以及质量的参差不齐,使得消费者获得有用的产品评论变得困难。为此,提出一种新的产品评论有用度评估方法。引入词向量作为评论文本的深度特征表示,结合结构特征、情感特征、元数据特征等训练回归模型,自动地对评论进行有用度评估并基于有用度对评论进行排序。在Amazon真实数据集上的实验结果表明,该方法在回归性能和排序性能上均优于UGR+LEN+STR方法和基准方法。另外通过挖掘特定领域的词向量特征,该向量模型在RMSE,NDCG等评价指标上可有效地改善评估效果。

关键词: 有用度, 深度学习, 神经网络语言模型

Abstract: Rapid increasing volume of product reviews as well as their variant qualities makes it time-consuming and tedious for individuals to perceive valuable reviews.Therefore,this paper proposes an approach to automatically assess helpfulness of online product reviews based on word vector.The approach introduces word vector as a deep text feature and incorporates it with structure feature,sentiment feature and meta feature to learn a regression model for automatically helpfulness assessing.A ranking procedure based on helpfulness is performed.Compared with UGR+LEN+STR’s approach and the baseline,experimental results conducted on Amazons dataset show that this approach achieves promising performances both on regression and ranking.Furthermore,this paper explores domain-specific word vector model,which can improve assessing effect on RMSE,NDCG and other evaluation indexs.

Key words: product reviews, word vector, helpfulness, deep learning, neural network language model

中图分类号: 

• TP18