Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

QAEA-DR: A Unified Text Augmentation Framework for Dense Retrieval

Hongming Tan, Shaoxiong Zhan, Hai Lin, Hai-Tao Zheng, Wai Kin Chan

TL;DR

QAEA-DR introduces a unified text augmentation framework that enhances dense retrieval by generating information-dense QA pairs and element-driven events from raw texts using LLM prompting. The approach integrates these QA and EE representations into the vector database via two organization modes (TRI and TMO) and a scoring/regeneration loop, improving retrieval fidelity without changing embedding or retrieval methods. Theoretical analysis shows fidelity improvements under realism conditions, while extensive experiments across multilingual datasets and multiple LLMs demonstrate robust, scalable gains in NDCG and retrieval quality. The work highlights practical trade-offs between generation quality, diversity, dataset size, and computational cost, positioning QAEA-DR as a versatile open-domain text optimizer for retrieval-augmented generation systems.

Abstract

In dense retrieval, embedding long texts into dense vectors can result in information loss, leading to inaccurate query-text matching. Additionally, low-quality texts with excessive noise or sparse key information are unlikely to align well with relevant queries. Recent studies mainly focus on improving the sentence embedding model or retrieval process. In this work, we introduce a novel text augmentation framework for dense retrieval. This framework transforms raw documents into information-dense text formats, which supplement the original texts to effectively address the aforementioned issues without modifying embedding or retrieval methodologies. Two text representations are generated via large language models (LLMs) zero-shot prompting: question-answer pairs and element-driven events. We term this approach QAEA-DR: unifying question-answer generation and event extraction in a text augmentation framework for dense retrieval. To further enhance the quality of generated texts, a scoring-based evaluation and regeneration mechanism is introduced in LLM prompting. Our QAEA-DR model has a positive impact on dense retrieval, supported by both theoretical analysis and empirical experiments.

QAEA-DR: A Unified Text Augmentation Framework for Dense Retrieval

TL;DR

QAEA-DR introduces a unified text augmentation framework that enhances dense retrieval by generating information-dense QA pairs and element-driven events from raw texts using LLM prompting. The approach integrates these QA and EE representations into the vector database via two organization modes (TRI and TMO) and a scoring/regeneration loop, improving retrieval fidelity without changing embedding or retrieval methods. Theoretical analysis shows fidelity improvements under realism conditions, while extensive experiments across multilingual datasets and multiple LLMs demonstrate robust, scalable gains in NDCG and retrieval quality. The work highlights practical trade-offs between generation quality, diversity, dataset size, and computational cost, positioning QAEA-DR as a versatile open-domain text optimizer for retrieval-augmented generation systems.

Abstract

In dense retrieval, embedding long texts into dense vectors can result in information loss, leading to inaccurate query-text matching. Additionally, low-quality texts with excessive noise or sparse key information are unlikely to align well with relevant queries. Recent studies mainly focus on improving the sentence embedding model or retrieval process. In this work, we introduce a novel text augmentation framework for dense retrieval. This framework transforms raw documents into information-dense text formats, which supplement the original texts to effectively address the aforementioned issues without modifying embedding or retrieval methodologies. Two text representations are generated via large language models (LLMs) zero-shot prompting: question-answer pairs and element-driven events. We term this approach QAEA-DR: unifying question-answer generation and event extraction in a text augmentation framework for dense retrieval. To further enhance the quality of generated texts, a scoring-based evaluation and regeneration mechanism is introduced in LLM prompting. Our QAEA-DR model has a positive impact on dense retrieval, supported by both theoretical analysis and empirical experiments.
Paper Structure (28 sections, 2 theorems, 14 equations, 6 figures, 12 tables, 1 algorithm)

This paper contains 28 sections, 2 theorems, 14 equations, 6 figures, 12 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Dense Retrieval
  4. Input Enhancement in Retrieval
  5. Information Extraction
  6. Approach
  7. Notations and Problem Definition
  8. Overview of QAEA-DR
  9. LLM-based Text Augmentation in QAEA
  10. Theory in QAEA-DR
  11. Experiment
  12. Datasets and Baselines
  13. Implementation Details
  14. Evaluation Metrics
  15. Main Results
  16. ...and 13 more sections

Key Result

Theorem 3.3

Given a text $t_i$, let $\{v_i^{(j)}\}$ represent a set of generated text vectors, where $j$ records the total number of generated texts, and let $v_i$ represent the original text vector. Consider a text $t_1$ most relevant to a query $q$ and a competing text $t_2$, we have generated text vector set and are both satisfied under the following conditions: (i) Relevance Enhancement: $\exists v_1^{(0

Figures (6)

  • Figure 1: QAEA-DR example. The dashed arrows represent the text augmentation path of QAEA. In Step-1: High-level Structured Information Extraction through LLM-based text generators with frozen parameters, the generated QA pairs and events preserve similar key information in different formats. This is followed by Step-2: Reversion to Unstructured Form and Step-3: Integration into the Vector Database. In Step-4: Enhanced Dense Retrieval, the results demonstrate that there exists a generated vector with higher query relevance compared to the original text vector. This is because the key information density of the generated text vector is enhanced through information extraction, making it semantically closer to the query.
  • Figure 2: Prompt templates for QA pair/event generation, scoring-based quality evaluation and regeneration, respectively.
  • Figure 3: QAEA-DR vs Baseline on Recall@1. The blue bar represents the percentage of the entire dataset where QAEA-DR correctly recalls at rank 1, while the baseline does not; the orange bar indicates the opposite. The difference between the blue and orange bars quantifies the actual improvement that QAEA-DR provides over the baseline.
  • Figure 4: Ablation Performance of QAEA: NDCG@1 ($\times 100$)
  • Figure 5: Analysis of different dataset sizes: NDCG@1 ($\times 100$)
  • ...and 1 more figures

Theorems & Definitions (6)

  • Definition 3.1: QAEA-DR
  • Definition 3.2: Normalized Margin
  • Theorem 3.3: Text Augmentation
  • proof
  • Theorem 3.4: QAEA
  • proof