Breaking the Top-k Assumption in Pseudo-Relevance Feedback: An Empirical Analysis of Relevance Distribution in BM25 Rankings

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Pseudo-Relevance Feedback (PRF) has remained a cornerstone of unsupervised retrieval since Rocchio (1971), yet the foundational assumption that the top- retrieved documents are the best available feedback has received limited direct empirical scrutiny, despite widespread adoption in both classical and neural approaches. This study analyzed BM25 retrieval on the TREC Deep Learning 2019 and 2020 test collections (97 queries, 20,646 graded relevance judgements) and found that 79% of relevant documents fall outside the top-10, with a mean rank of 37.3. An oracle selection strategy achieved 0.324 higher feedback precision at , defined as the proportion of graded-relevant documents within the selected for expansion, with a large effect size that is consistent across all tested values of and both test collections. LLM-based analysis of 61 extreme cases identified vocabulary gap as the dominant failure mode in 96.7% of cases, driven primarily by implicit relevance (35.6%) and hypernym-hyponym mismatch (27.1%). These findings establish that document selection, rather than term weighting, is the primary lever for PRF improvement and identify the vocabulary gap as the principal target for next-generation methods. The results demonstrate that improving feedback-document selection represents a largely unexplored avenue for PRF advancement.