Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Dense Passage Retrieval: Is it Retrieving?

Benjamin Reichman, Larry Heck

TL;DR

The paper investigates how dense passage retrieval (DPR) fine-tuning affects knowledge representation in transformer models within the RAG framework. Using probing, layer-activation, and model-editing analyses on a BERT backbone, it finds that DPR decentralizes how knowledge is stored and accessed without adding new facts, and that removal of facts from the base model remains unrecovered by DPR unless the edits are incorporated. The work highlights multiple retrieval pathways and shows that DPR relies on pre-existing knowledge, suggesting future directions to decentralize knowledge further and to map internal knowledge to external knowledge bases. These insights have practical implications for reducing hallucinations and improving the precision and robustness of retrieval-augmented systems in real-world applications.

Abstract

Dense passage retrieval (DPR) is the first step in the retrieval augmented generation (RAG) paradigm for improving the performance of large language models (LLM). DPR fine-tunes pre-trained networks to enhance the alignment of the embeddings between queries and relevant textual data. A deeper understanding of DPR fine-tuning will be required to fundamentally unlock the full potential of this approach. In this work, we explore DPR-trained models mechanistically by using a combination of probing, layer activation analysis, and model editing. Our experiments show that DPR training decentralizes how knowledge is stored in the network, creating multiple access pathways to the same information. We also uncover a limitation in this training style: the internal knowledge of the pre-trained model bounds what the retrieval model can retrieve. These findings suggest a few possible directions for dense retrieval: (1) expose the DPR training process to more knowledge so more can be decentralized, (2) inject facts as decentralized representations, (3) model and incorporate knowledge uncertainty in the retrieval process, and (4) directly map internal model knowledge to a knowledge base.

Dense Passage Retrieval: Is it Retrieving?

TL;DR

The paper investigates how dense passage retrieval (DPR) fine-tuning affects knowledge representation in transformer models within the RAG framework. Using probing, layer-activation, and model-editing analyses on a BERT backbone, it finds that DPR decentralizes how knowledge is stored and accessed without adding new facts, and that removal of facts from the base model remains unrecovered by DPR unless the edits are incorporated. The work highlights multiple retrieval pathways and shows that DPR relies on pre-existing knowledge, suggesting future directions to decentralize knowledge further and to map internal knowledge to external knowledge bases. These insights have practical implications for reducing hallucinations and improving the precision and robustness of retrieval-augmented systems in real-world applications.

Abstract

Dense passage retrieval (DPR) is the first step in the retrieval augmented generation (RAG) paradigm for improving the performance of large language models (LLM). DPR fine-tunes pre-trained networks to enhance the alignment of the embeddings between queries and relevant textual data. A deeper understanding of DPR fine-tuning will be required to fundamentally unlock the full potential of this approach. In this work, we explore DPR-trained models mechanistically by using a combination of probing, layer activation analysis, and model editing. Our experiments show that DPR training decentralizes how knowledge is stored in the network, creating multiple access pathways to the same information. We also uncover a limitation in this training style: the internal knowledge of the pre-trained model bounds what the retrieval model can retrieve. These findings suggest a few possible directions for dense retrieval: (1) expose the DPR training process to more knowledge so more can be decentralized, (2) inject facts as decentralized representations, (3) model and incorporate knowledge uncertainty in the retrieval process, and (4) directly map internal model knowledge to a knowledge base.
Paper Structure (12 sections, 2 equations, 7 figures, 3 tables)

This paper contains 12 sections, 2 equations, 7 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Knowledge Consistency Between Untrained and Trained Model
  3. Knowledge Decentralization in DPR-Trained Models
  4. Knowledge Editing Experiments
  5. Knowledge Removal
  6. Related Works
  7. Conclusion
  8. Limitations
  9. Ethics Statements
  10. Appendix
  11. Context Model Activations
  12. Model Activations at different thresholds

Figures (7)

  • Figure 1: Layerwise activations for pre-trained and DPR-trained BERT. The parenthetical numbers indicate the number of neurons in the layer that are above the attribution threshold for any number of examples.
  • Figure 2: Layerwise activations for pre-trained and DPR-trained BERT - context model. The parenthetical numbers indicate the number of neurons in the layer that are above the attribution threshold for any number of examples.
  • Figure 3: Layerwise activations for pre-trained and DPR-trained BERT with a threshold of 0.005. The parenthetical numbers indicate the number of neurons in the layer that are above the attribution threshold for any number of examples.
  • Figure 4: Layerwise activations for pre-trained and DPR-trained BERT with a threshold of 0.01. The parenthetical numbers indicate the number of neurons in the layer that are above the attribution threshold for any number of examples.
  • Figure 5: Layerwise activations for pre-trained and DPR-trained BERT with a threshold of 0.05. The parenthetical numbers indicate the number of neurons in the layer that are above the attribution threshold for any number of examples.
  • ...and 2 more figures