Integrating Video-Language Model with Mixture of Low-Rank Experts for Video Commonsense Captioning
TL;DR
A Video-Language Model with Mixture of Low-rank Adaptation Experts (VLM-MOLAE) that excels in understanding videos and performing commonsense reasoning, and utilizes a pre-trained large language model as the backbone to exploit its commonsense knowledge to facilitate comprehension and inference.
Abstract
The video commonsense captioning task aims to generate descriptive textual captions that encapsulate not only visible content, but also underlying events and their inherent commonsense. Unlike traditional video captions, which merely describe visible objects and actions, video commonsense captioning provides deeper insights. However, current research on video commonsense reasoning still faces challenges. Existing studies often focus on understanding explicit video content, and there is still a need to bridge the complex semantic gap between visible video content and intangible commonsense knowledge. Hence, we propose a novel and affordable solution. By leveraging the commonsense knowledge within large language models, we construct a Video-Language Model with Mixture Of Low-rank Adaptation Experts (VLM-MOLAE) that excels in understanding videos and performing commonsense reasoning. Specifically, we first utilize a pre-trained large language model as the backbone, which enables us to exploit its commonsense knowledge to facilitate comprehension and inference. Further, a transformer-based video embedder module is specially designed to extract video features to be aligned with the LLM input space. Additionally, a parameter-efficient mixture of experts module is developed to adapt the large language model with low-rank experts, enabling precise reasoning for diverse inputs. Extensive experiments and analyses conducted on a benchmark dataset demonstrate the effectiveness and superiority of our proposed model. The code is publicly available at https://github.com/gxytsyt/VLM_MOLAE.