Sentimental Image Generation for Aspect-based Sentiment Analysis

Recent research work on textual Aspect-Based Sentiment Analysis (ABSA) have achieved promising performance. However, a persistent challenge lies in the limited semantics derived from the raw data. To address this issue, researchers have explored enhancing textual ABSA with additional augmentations, they either craft audio (Guo et al., 2024), text (Seo et al., 2024) and linguistic features (Bao et al., 2022) based on the input, or rely on user-posted images (Yu and Jiang, 2019). Yet these approaches have their limitations: the former three formations are heavily overlap with the original data, which undermines their ability to be supplementary while the user-posted images are extremely dependent on human annotation, which not only limits its application scope to just a handful of text-image datasets, but also propagates the errors derived from human mistakes to the entire downstream loop. In this study, we explore the way of generating the sentimental image that no one has ever ventured before. We propose a novel Sentimental Image Generation method that can precisely provide ancillary visual semantics to reinforce the textual extraction as shown in Figure 1. Extensive experiments build a new SOTA performance in ACOS, ASQP and en-Phone datasets, underscoring the effectiveness of our method and highlighting a promising direction for expanding our features. Our code and data are available at https://github.com/ HoraceXIaoyiBao/ACL25-SIG4ABSA.