Research On Sentiment Analysis Based On ERNIE And Multi-feature Fusion

With the rapid development of social networks,people increasingly express their emotions on various social platforms regarding popular topics and new phenomena through textual information.These messages often include a large number of emoticons composed of icons or symbols,known as emojis,to convey emotional information.Analyzing the emotional intent of such information has become a valuable reference for governments,businesses,and marketers.Currently,leveraging deep learning to automatically process large-scale corpora and analyze sentiment polarity has become a focus of research.In sentiment analysis tasks,traditional Convolutional Neural Networks(CNNs)can extract local features using convolutional kernels but often overlook long-range contextual information.On the other hand,Bidirectional Long Short-Term Memory Networks(Bi LSTMs)can capture contextual dependencies but have relatively weaker local feature extraction capabilities.Additionally,existing sentiment analysis tasks typically choose to ignore emoji symbol features,leading to a certain degree of loss in sentiment features.To address these issues,this study proposes the LE-CNN-MBi LSTM model and the LE-DECNN-MBi LSTM model to enhance sentiment analysis performance.The LE-CNN-MBi LSTM model introduces parallel CNNs and dual-channel Bi LSTMs.It utilizes CNNs to extract multiple local key features from the text and employs Bi LSTMs to extract contextual semantics.The CNN-Bi LSTM module is utilized to extract fused features.Moreover,the text embedding process incorporates text category labels and concatenates the hierarchical ERNIE model layers to form a richer text vector representation.The LE-DECNN-MBi LSTM model,based on the former,includes an emoji symbol processing module that addresses the ambiguity of emoji symbols through weighted dual-embedding.It also introduces self-attention mechanism by feeding the features from multiple models into it,enabling higher attention to important information in the text and achieving higher accuracy,F1 score,and recall rate.By conducting comparative experiments with other models,the results demonstrate that the proposed models perform well in sentiment analysis tasks,with the LE-DECNN-MBi LSTM model exhibiting outstanding performance in analyzing the sentiment of texts containing emoji symbols.Finally,based on these models,a sentiment analysis testing system is built,which can analyze user-inputted texts and provide corresponding sentiment tendencies,helping users accurately identify the sentiment polarity of comments.