📄 Enhancing Retrieval via Cognitively Motivated Document Expansion

Summary

The paper asks whether LLMs can strengthen first-stage document retrieval by rewriting the corpus rather than the query, an approach the authors note has received less attention than query expansion. Taking inspiration from the 5E Instructional Model, a framework rooted in educational psychology, they wrote six prompts grouped into three macro-categories (EN, EX, EL), leaving out the Evaluate phase. The prompts ask an LLM, acting as a "student," to summarise, enrich with its own knowledge, or reframe each document independently of any query. Every prompt yields a separate generated corpus. Queries are run against each variant, and the resulting rankings are combined with fusion algorithms. Tests used Qwen2.5 Instruct models and embedding-based retrieval on 13 retrieval datasets, scored with nDCG@10. No single generated corpus beat the original. However, fusing rankings from several generated corpora, especially together with the original, frequently improved results. The authors conclude that LLM rewrites act as useful complementary views of a collection rather than replacements for it.

Key points

• The prompt set has six prompts, two per macro-category:
  • EN: summarising and re-wording with more salient vocabulary.
  • EX: knowledge-based expansion, and condensing while adding knowledge.
  • EL: a negative-stance rewrite and an explanatory commentary.
• Prompts that summarise or add LLM knowledge (EN, EX) helped more than unusual reformulations (EL).
• Three Qwen2.5 Instruct LLMs were used. Comparing Qwen2.5-7B-Instruct and Qwen2.5-14B-Instruct showed no consistent advantage for the larger model.
• Main experiments used Infly embeddings. With the weaker gte-Qwen2-1.5B-instruct, expansion helped more, and fusion without the original slightly beat fusion with it.
• The authors flag several limitations:
  • No significance tests were run.
  • Baselines deviate slightly from MTEB leaderboard scores.
  • Code is available on GitHub.