A multi-label text sentiment analysis model based on sentiment correlation modeling

Objective: This study proposes an emotion correlation-enhanced sentiment analysis model (ECO-SAM), a sentiment correlation modeling-based multi-label sentiment analysis model.

Methods: The ECO-SAM utilizes a pre-trained BERT encoder to obtain semantic embedding of input texts and then leverages a self-attention mechanism to model the semantic correlation between emotions. Additionally, it utilizes a text emotion matching neural network to make sentiment analysis for input texts.

Results: The experiment results in public datasets demonstrate that compared to baseline models, the ECO-SAM obtains the precision score increasing by 13.33% at most, the recall score increasing by 3.69% at most, and the F1 score increasing by 8.44% at most. Meanwhile, the modeled sentiment semantics are interpretable.

Limitations: The data modeled by the ECO-SAM are limited to text-only modality, excluding multi-modal data that could enhance classification performance. Additionally, the training data are not large-scale, and there is a lack of high-quality large-scale training data for fine-tuning sentiment analysis models.

Conclusion: The ECO-SAM is capable of effectively modeling sentiment semantics and achieving excellent classification performance in many public sentiment analysis datasets.