Improving Interpretability in Document-Level Polarity Classification by Applying Attention

Document-level polarity classification has attracted interest in the real world. While LLMs have made it possible for accurate classification, these complex models have the problem of interpretability. Our contribution is to apply inter-sentence attention, which captures the relationship between sentences, to a more practical interpretable model. By utilizing high inter-sentence attention scores, meaning corresponding sentences are related to each other, we attempt to capture the context of sentences and make them more similar to the human judgment process. With two real datasets, we compared our model with prior models in terms of classification performance and interpretability and found that our model is more accurate on both datasets. In addition, to assess interpretability, we examined the overlap between sentences that contribute to the model's predictions and those annotated by humans for the same document. The results show that our model has a larger overlap and is more likely to extract interpretive sentences that humans intuitively consider important. In addition, our result partially captures the polarity of 'implicit' sentences that do not contain direct expressions, which could not be captured by prior models, suggesting that our model may lead to a more natural interpretation.