A guard against ambiguous sentiment for multimodal aspect-level sentiment classification

Recent advances in multimodal learning have achieved state-of-the-art results in aspect-level sentiment classification by leveraging both text and image data. However, images can sometimes contain contradictory sentiment cues or convey complex messages, making it difficult to accurately determine the sentiment expressed in the text. Intuitively, we should only use image data to complement the text if the latter contains ambiguous sentiment or leans toward the neutral polarity. Therefore, instead of trying to forcefully use images as done in prior work, we develop a Guard against Ambiguous Sentiment (GAS) for multimodal aspect-level sentiment classification (MALSC). Built on a pretrained language model, GAS is equipped with a novel “ambiguity learning” strategy that focuses on learning the degree of sentiment ambiguity within the input text. The sentiment ambiguity then serves to determine the extent to which image information should be utilized for accurate sentiment classification. In our experiments with two benchmark twitter datasets, we found that GAS achieves a performance gain of up to 0.98% in macro-F1 score compared to recent methods in the task. Furthermore, we explore the efficacy of large language models (LLMs) in the MALSC task by employing the core ideas behind GAS to design tailored prompts. We show that multimodal LLMs such as LLaVA, when provided with GAS-principled prompts, yields a 2.4% improvement in macro-F1 score for few-shot learning on the MALSC task.