Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Generative Multi-Modal Knowledge Retrieval with Large Language Models

Xinwei Long, Jiali Zeng, Fandong Meng, Zhiyuan Ma, Kaiyan Zhang, Bowen Zhou, Jie Zhou

TL;DR

This paper tackles knowledge retrieval in multi-modal settings by reframing it as an end-to-end generative task. It introduces GeMKR, which uses LLMs as virtual knowledge bases and retrieves knowledge through a two-step process: generate informative knowledge clues and then query a database with those clues, guided by object-aware prefix tuning and multi-modal alignment. A knowledge-guided constrained decoding strategy ensures clues map uniquely to documents, enabling efficient and scalable retrieval across large KBs while requiring relatively little training data. Empirical results on three benchmarks show substantial gains over strong baselines, demonstrating the method’s effectiveness and scalability for real-world multi-modal knowledge-intensive applications.

Abstract

Knowledge retrieval with multi-modal queries plays a crucial role in supporting knowledge-intensive multi-modal applications. However, existing methods face challenges in terms of their effectiveness and training efficiency, especially when it comes to training and integrating multiple retrievers to handle multi-modal queries. In this paper, we propose an innovative end-to-end generative framework for multi-modal knowledge retrieval. Our framework takes advantage of the fact that large language models (LLMs) can effectively serve as virtual knowledge bases, even when trained with limited data. We retrieve knowledge via a two-step process: 1) generating knowledge clues related to the queries, and 2) obtaining the relevant document by searching databases using the knowledge clue. In particular, we first introduce an object-aware prefix-tuning technique to guide multi-grained visual learning. Then, we align multi-grained visual features into the textual feature space of the LLM, employing the LLM to capture cross-modal interactions. Subsequently, we construct instruction data with a unified format for model training. Finally, we propose the knowledge-guided generation strategy to impose prior constraints in the decoding steps, thereby promoting the generation of distinctive knowledge clues. Through experiments conducted on three benchmarks, we demonstrate significant improvements ranging from 3.0% to 14.6% across all evaluation metrics when compared to strong baselines.

Generative Multi-Modal Knowledge Retrieval with Large Language Models

TL;DR

This paper tackles knowledge retrieval in multi-modal settings by reframing it as an end-to-end generative task. It introduces GeMKR, which uses LLMs as virtual knowledge bases and retrieves knowledge through a two-step process: generate informative knowledge clues and then query a database with those clues, guided by object-aware prefix tuning and multi-modal alignment. A knowledge-guided constrained decoding strategy ensures clues map uniquely to documents, enabling efficient and scalable retrieval across large KBs while requiring relatively little training data. Empirical results on three benchmarks show substantial gains over strong baselines, demonstrating the method’s effectiveness and scalability for real-world multi-modal knowledge-intensive applications.

Abstract

Knowledge retrieval with multi-modal queries plays a crucial role in supporting knowledge-intensive multi-modal applications. However, existing methods face challenges in terms of their effectiveness and training efficiency, especially when it comes to training and integrating multiple retrievers to handle multi-modal queries. In this paper, we propose an innovative end-to-end generative framework for multi-modal knowledge retrieval. Our framework takes advantage of the fact that large language models (LLMs) can effectively serve as virtual knowledge bases, even when trained with limited data. We retrieve knowledge via a two-step process: 1) generating knowledge clues related to the queries, and 2) obtaining the relevant document by searching databases using the knowledge clue. In particular, we first introduce an object-aware prefix-tuning technique to guide multi-grained visual learning. Then, we align multi-grained visual features into the textual feature space of the LLM, employing the LLM to capture cross-modal interactions. Subsequently, we construct instruction data with a unified format for model training. Finally, we propose the knowledge-guided generation strategy to impose prior constraints in the decoding steps, thereby promoting the generation of distinctive knowledge clues. Through experiments conducted on three benchmarks, we demonstrate significant improvements ranging from 3.0% to 14.6% across all evaluation metrics when compared to strong baselines.
Paper Structure (27 sections, 8 equations, 4 figures, 5 tables)

This paper contains 27 sections, 8 equations, 4 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Multi-Modal Knowledge Retrieval
  4. Generative Retrieval
  5. Large Language Model and Efficient Fine-tuning
  6. Methodology
  7. Problem Formulation
  8. Object-aware Prefix Tuning
  9. Multi-Modal Alignment
  10. Instruction Tuning
  11. Supervised Text Sampling.
  12. Instruction Data Construction.
  13. Model Training.
  14. Knowledge-guided Constraint Decoding
  15. Experiments
  16. ...and 12 more sections

Figures (4)

  • Figure 1: Multi-Modal Knowledge Retrieval. Prior studies use multiple retrievers for separate purposes, while we retrieve knowledge through an end-to-end generative model.
  • Figure 2: Overall architecture of our GeMKR.
  • Figure 3: Results of scaling up the LLMs
  • Figure 4: Case Study. Three cases from the OKVQA-GS112K dataset. Each predicted knowledge clue can be uniquely mapped to a document in the KB. The predicted knowledge clues that occur in corresponding documents are highlighted in yellow.