摘要：

随着手机应用软件的流行，应用市场上出现了大量非结构化的中文用户评论。基于用户评论识别App用户意图，可以帮助开发人员对App软件进行有针对性的维护和改善。为了从中准确识别用户意图，提出一种基于融合模型和语义网络的App用户意图识别方法FSAUIR。使用百度工具Senta判断评论的情感倾向，构建基于RoBERTa的融合意图分类模型RBMS，通过RoBERTa模型将用户评论转化为语义特征表示，并将其输入到双向门控循环单元中，以提取评论的全局上下文语义信息，同时利用多头自注意力机制和SoftPool获取关键的特征信息，保留主要特征，通过Softmax进行归一化处理，得到意图分类结果。在意图分类的基础上，引入PositionRank模型提取各意图类别下评论的关键词，计算关键词之间的共现关系，构建关键词语义网络，从而更细粒度地识别用户意图。实验结果表明，相比BERT、RoBERTa、RoBERTa-CNN等模型，RBMS模型在人工标注数据集上具有较优的分类性能，准确率、精确率、召回率、F1值分别为87.75%、88.09%、87.80%、87.88%。此外，在意图分类的结果集中，FSAUIR构建的语义网络可以高效地挖掘出用户评论中有价值的信息。

关键词: 意图识别, 意图分类, RoBERTa模型, 双向循环门控单元, PositionRank模型, 多头自注意力机制

Abstract:

With the popularity of mobile Applications(Apps), a large number of unstructured Chinese user reviews have appeared in the application market. Identifying App user intent based on these reviews helps developers make targeted maintenance and improvement of App software. To accurately recognize user intent, this study proposes an App user intent recognition method based on fusion model and semantic network, named FSAUIR. First, FSAUIR uses the Baidu tool Senta to determine the emotional tendency of the reviews. It then introduces Robustly optimized Bidirectional Encoder Representation from Transformers approach(RoBERTa)-based fusion intent classification model, RoBERTa-BiGRU-Multiple Self-Attention+SoftPool(RBMS), which transforms user reviews into semantic feature representations through the RoBERTa model. These representations are input into a Bidirectional Gated Recurrent Unit(BiGRU) to extract the global contextual semantic information of the reviews. Simultaneously, the multiple self-attention and SoftPool mechanisms obtain more critical feature information, retaining the main features. Finally, the Softmax normalizes the features to obtain the intent classification results. Subsequently, FSAUIR employs the PositionRank model to extract keywords from reviews under each intent category, calculate the co-occurrence relationship between keywords, and construct a keywords semantic network to recognize user intent with finer granularity. Experimental results show that compared to BERT, RoBERTa, RoBERTa-CNN, and other models, the RBMS model exhibits superior classification performance on the manually labeled dataset. The model achieves accuracy, precision, recall, and F1 value of 87.75%, 88.09%, 87.80%, and 87.88%, respectively. Additionally, the semantic network constructed by FSAUIR efficiently mines valuable information from user reviews in the intent classification result set.

Key words: intent recognition, intent classification, RoBERTa model, Bidirectional Gated Recurrent Unit(BiGRU), PositionRank model, multihead self-attention mechanism