AI Research

Multi-Granularity Context-Enhanced RAG over Multimo... | AI Research

Key Takeaways

The paper "Multi-Granularity Context-Enhanced RAG over Multimodal Knowledge Graphs" introduces a framework called Context-Enhanced Multimodal Knowledge Graph...
Retrieval-augmented generation (RAG) is widely used to mitigate hallucination issues in large language models (LLMs) and multimodal large language models (MLLMs).
In particular, knowledge graph (KG)-based RAG leverages structured knowledge to provide (M)LLMs with high-quality external information.
Building on these works, recent studies have explored multimodal knowledge graphs (MMKGs) as knowledge bases for GraphRAG.
This enables Graph RAG to integrate knowledge across multiple modalities, thereby further enhancing its performance.

Paper AbstractExpand

Retrieval-augmented generation (RAG) is widely used to mitigate hallucination issues in large language models (LLMs) and multimodal large language models (MLLMs). In particular, knowledge graph (KG)-based RAG leverages structured knowledge to provide (M)LLMs with high-quality external information. Building on these works, recent studies have explored multimodal knowledge graphs (MMKGs) as knowledge bases for GraphRAG. This enables Graph RAG to integrate knowledge across multiple modalities, thereby further enhancing its performance. However, existing MMKG-based RAG methods generally follow a common pipeline in which different modalities are largely processed independently before being fusion. As a result, textual context is only used to a limited extent during visual information extraction and subsequent multimodal knowledge fusion. This brings a semantic gap between images and text which limits the multimodal GraphRAG performance. To address this issue, we propose a novel framework for constructing a Context-Enhanced MMKG (CEMMKG) to better support multimodal GraphRAG. The proposed CEMMKG enriches each image with complementary textual context at both local and global scopes. Local context goes beyond the surrounding text by incorporating sentences that are semantically related to the image, while global context provides a summary of the entire passage. We further introduce a multi-granularity design for the local context, allowing it to capture semantically relevant information at different levels of detail. Extensive experiments on the selected vision-centric dataset validate that CEMMKG is effective in leveraging contextual information to improve MMKG-based RAG performance. Moreover, its effectiveness across different MMKG-based RAG methods demonstrates its broad applicability.

The paper "Multi-Granularity Context-Enhanced RAG over Multimodal Knowledge Graphs" introduces a framework called Context-Enhanced Multimodal Knowledge Graph (CEMMKG) to improve how multimodal large language models (MLLMs) retrieve and use information. The authors address a common issue in current systems where visual and textual data are processed separately, creating a "semantic gap" that limits the effectiveness of Retrieval-Augmented Generation (RAG).

Addressing the Modality Gap

Current multimodal RAG methods typically process images and text independently before merging them into a knowledge graph. Because these systems often rely only on text immediately surrounding an image, they frequently miss important information distributed elsewhere in a document. The authors argue that this limited approach results in suboptimal visual knowledge extraction and weak connections between images and text. CEMMKG aims to solve this by systematically incorporating relevant textual context into the construction of the knowledge graph.

How the Framework Works

The CEMMKG framework enriches visual elements with two types of textual context:

Local Context: This captures fine-grained details by incorporating sentences that are semantically related to a specific image, moving beyond simple proximity. The framework uses a multi-granularity design, allowing it to capture information at different levels of detail.
Global Context: This provides a high-level summary of the entire document to ensure the model has a holistic understanding of the content.
These contexts are integrated during the construction of the multimodal knowledge graph, specifically supporting the stages where visual information is converted into graph nodes and where visual and textual data are fused together.

Experimental Results

The researchers tested the CEMMKG framework on a vision-centric subset of the MMLongBench-Doc dataset. According to the paper, the experiments validate that incorporating multi-granular contextual information improves the performance of MMKG-based RAG. The authors also note that the framework is designed to be compatible with various existing MMKG-based RAG methods, suggesting it has broad applicability for improving how these systems handle complex, document-based information.

Key Considerations

The authors identify that the quality of a knowledge graph is fundamental to the accuracy of the final answers generated by an MLLM. By focusing on the construction phase of the graph—rather than just the retrieval phase—the researchers aim to ensure that the underlying data structure is as comprehensive as possible. The study emphasizes that simply adding more text is not enough; instead, the design and selection of context must be carefully aligned with the visual elements to bridge the gap between different data modalities.

Comments (0)

No comments yet

Be the first to share your thoughts!