基于神经网络模型的产品属性情感分析

发布时间：2019-05-10 03:41

【摘要】：针对基于词向量的神经网络模型在产品属性情感分析中效果不佳的问题,提出一种集成离散特征和词向量特征的开关递归神经网络模型。首先,通过直接循环图为语句建模,采用开关递归神经网络模型完成产品属性情感分析任务;然后,在开关递归神经网络模型中集成离散特征和词向量特征;最后,分别在流水线、联合、折叠三种任务模型中完成属性提取和情感分析任务。以宏观F1分数作为评估指标,在Sem Eval-2014的笔记本电脑和餐馆评论数据集上做实验。开关递归神经网络模型的F1分数为:48.21%和62.19%,超过普通递归神经网络模型近1.5个百分点,因而开关递归神经网络能够有效捕获复杂特征,提升产品属性情感分析的效果。而集成离散特征和词向量特征的神经网络模型的F1分数为:49.26%和63.31%,均超过基线结果 0.5到1个百分点,表明离散特征和词向量特征互相促进,另一方面,也表明仅仅基于词向量的神经网络模型仍有提升空间。三种任务模型中,流水线模型的F1分数最高,表明应将属性提取和情感分析任务分开完成。
[Abstract]:In order to solve the problem that the neural network model based on word vector is not effective in product attribute emotion analysis, a switching recurrent neural network model integrating discrete features and word vector features is proposed. Firstly, the sentence is modeled by direct loop graph, and the switching recurrent neural network model is used to complete the task of product attribute emotion analysis, and then, the discrete features and word vector features are integrated into the switching recurrent neural network model. Finally, attribute extraction and emotion analysis are completed in pipeline, union and folding task models. Using macro F1 score as an evaluation index, experiments were carried out on Sem Eval-2014 laptops and restaurant review datasets. The F1 scores of switching recurrent neural network model are 48.21% and 62.19%, which are nearly 1.5 percentage points higher than those of ordinary recurrent neural network model, so the switching recurrent neural network can effectively capture complex features. Improve the effect of emotional analysis of product attributes. The F1 scores of the neural network model integrating discrete features and word vector features are 49.26% and 63.31%, both of which exceed the base line results by 0.5 to 1 percentage point, indicating that discrete features and word vector features promote each other, on the other hand, It is also shown that there is still room for lifting in the neural network model based only on word vectors. Among the three task models, the F1 score of pipeline model is the highest, which indicates that attribute extraction and emotion analysis should be completed separately.
【作者单位】：武汉大学计算机学院;
【基金】：国家自然科学基金资助项目(61133012)~~
【分类号】：TP183;TP391.1

【相似文献】