Exploring Multi-Model Sentiment Analysis: A Conditional-Value System

This paper presents a detailed comparative study conducted on five distinctive language models with the objective to classify sentiments embedded in a dataset of 10,000 reviews. Each model allocated sentiment scores of −1, 0, or 1, representing negative, neutral, or positive sentiments respectively. The intricacies of the study became evident when varying models assigned divergent classifications. In response to these discrepancies, we introduced a rule-based methodology, assigning a ‘Conditional-Value’ to each data entry, with an average score of 0.3326 across the dataset. Our study's results underscored the dependability of the ‘Cardiffnlp-label’ model, particularly in instances of non-consensus among models. The rule-based process, employed to ascertain ‘Conditional-Value’, offers a robust groundwork for subsequent studies in the realm of multi-model sentiment analysis. Moreover, the findings of this study underscore the potential advantages of developing a composite model, leveraging the unique strengths of varied models to augment precision in sentiment categorization. Despite the inherent intricacies entwined in sentiment analysis, this scholarly inquiry highlights the encouraging avenues for subsequent investigations and progress in this continually shifting and advancing discipline.