Refined Word Embeddings with Intensity Awareness for Fine-Level Sentiment Classification

Sentiment analysis employs classification models to discern people's opinions automatically. Recent strides made with Large Language Models (LLMs) have significantly enhanced the accuracy of binary-level sentiment classification, particularly through zero-shot and few-shot learning approaches. However, when it comes to fine-level sentiment classification, LLMs face challenges because they are not specifically trained for this downstream task. In contrast, other classification models utilize word embedding, a vector representation of words, as input data. Contemporary word embedding algorithms create these embeddings by considering the surrounding context of each word. Nonetheless, these word embeddings often fail to capture the nuances of intensity differences between words. For example, words like 'more' and 'less' have embeddings closely positioned in the semantic space, despite representing distinct intensity levels. These intensity words such as 'much', 'more', and 'less' are frequently used to convey the strength of opinions. Their intensity distinctions are crucial in fine-level sentiment classification. This paper introduces an innovative intensity-aware feed-forward neural network, equipped with a novel referential loss function designed to capture these intensity differences between words. The proposed model effectively separates words of varying intensities while bringing together words sharing the same intensity in the semantic space. To assess the effectiveness of this refined word embedding in sentiment analysis tasks, diverse fine-level sentiment datasets are employed. The results demonstrate that the refined word embedding surpasses original embeddings and popular LLMs for fine-level sentiment analysis