تحليل احساسات چندماهيتي مبتني برجنبه با غني‌سازي نمايش داده‌ها

Sentiment analysis, especially in the dynamic environment of social networks, is crucial for understan‎ding user opinions an‎d reactions. Among its variants, Aspect-Based Sentiment Analysis (ABSA) has gained attention as a multi-task problem aiming to simultaneously identify specific aspects discussed in a text an‎d determine the precise sentiment associated with each. The use of multimodal content (text an‎d image) to aid sentiment analysis poses a significant challenge, however, due to the inability of existing models to comprehensively extract information from both modalities an‎d create an effective semantic alignment for the two concurrent tasks (aspect extraction an‎d sentiment classification). This research focuses on the concept of targeted data representation enrichment an‎d proposes two novel, integrated frameworks to address these challenges. The core innovation lies in a dual, task-focused enrichment strategy: the primary approach centers on the comprehensive an‎d purposeful utilization of image affective information, an‎d the secondary approach is based on the structural analysis of the text. The primary approach enriches the visual representation by leveraging the affective an‎d aesthetic information of the image to significantly enhance the sentiment classification performance of the extracted aspects. To implement this, the proposed framework, Emotelligence-Adv, utilizes the capability of vision-language models to generate rich affective descriptions from the image an‎d text, thus creating a powerful an‎d targeted representation of the multimodal content. Furthermore, to ensure the extracted features are guided towards affective an‎d aesthetic components, an auxiliary task is dedicated to the direct prediction of image affect an‎d aesthetics. Parallel to this, to enrich the textual representation an‎d improve the aspect extraction task, graph structures an‎d textual dependency modeling are used to enhance the level of textual features. These approaches are integrated into a single framework, achieving a synergistic approach. Experimental results on the Twitter-2015 an‎d Twitter-2017 datasets demonstrate that the two proposed frameworks improve performance over the baseline model. Specifically, the Emotelligence-Adv approach coupled with the text graph module increased the F1 score by approximately 1.88% an‎d 1.73% compared to the baseline model, achieving the best performance on both datasets. Significantly, this model, by utilizing powerful vision-language models, is designed with a smaller number of trainable parameters than other models examined, highlighting its computational efficiency. This research demonstrates that the intelligent integration of affective an‎d structural aspects is an effective solution for more accurate an‎d robust sentiment analysis in complex multimodal environments.

6

دكتر عليرضا درويشي