Context-aware chatbot using MLLMs for Cultural Heritage

Multi-modal Large Language Models (MLLMs) are currently an extremely active research topic for the multimedia and computer vision communities, and show a significant impact in visual analysis and text generation tasks. MLLM's are well-versed in integrated understanding, analysis of complex data from cross modalities (i.e. text-image) and text generation with chat abilities. Almost all MLLM's, focus on alignment of image features to textual features for downstream text generation tasks includes detailed image description, visual question answering, stories and poems generation, phrase grounding, etc.. However, when focusing on visual question answering, questions that are highly relevant to the context of an image may not be answered correctly with the existing MLLM's, contrary to questions that are related to visual aspects. Moreover, generating meta data (context) for an image using present day MLLM's is hard task due to hallucinating characteristic of underlying Large Language Models (LLM's), and adequate contextual information cannot be directly derived from an image based perspective.