Prompt-Guided Distillation from Multimodal Large Language Models to Task-specific Models for Multimodal Sentiment Analysis
Keywords: Multimodal Sentimen Analysis, Representation Learning, Multimodal Large Language Model, Knowledge Distillation
Abstract: Multimodal Sentiment Analysis (MSA) has made some progress with the advent of Multimodal Large Language Models (MLLMs). However, the scalability and the closed-source nature of some MLLMs imposes challenges for efficient application in the real-word. In this study, we explore an innovative pathway to infuse the capabilities of general MLLMs into task-specific small models for MSA. We introduce the Prompt-Guided Multimodal Framework (PGMF), a refined teacher-student framework designed to transfer knowledge from powerful, general MLLMs to smaller, efficient models. The PGMF-Teacher utilizes MLLM-generated prompts and a tailored conditional alignment module to achieve better MSA, while the PGMF-Student distills this expertise to predict independently of MLLMs' guidance. Extensive evaluations on two popular MSA datasets including SIMS and MOSI demonstrate that compared to previous task-specific small models, PGMF-Teacher achieves state-of-the-art performance with the help of MLLMs' prompts, while PGMF-Student achieve competitive results with fewer parameters and without relying on MLLMs' prompts. The proposed framework offers a novel way to equip task-specific small models with the capability of MLLMs.
Primary Area: applications to computer vision, audio, language, and other modalities
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 379
Loading