Concetti Chiave
Coarse-tuning bridges pre-training and fine-tuning, improving document retrieval effectiveness.
Sintesi
Coarse-tuning introduced as an intermediate learning stage.
Query representations and query-document relations learned in coarse-tuning.
Evaluation experiments show significant improvements in MRR and nDCG@5.
ORCAS dataset used for training data.
Proposed method outperforms fine-tuned baseline in various datasets.
Statistiche
"Evaluation on Robust04: the symbols (*, †) indicates that there was a significant difference (p<0.01, p<0.05, respectively) compared to fine-tuned (baseline)."
"Table 1: Evaluation on Robust04: the most effective method was coarse+fine (proposed)."
"Table 2: Evaluation on GOV2, TREC-COVID, and TREC-DL: coarse+fine (proposed) outperformed fine-tuned (baseline) in MRR and nDCG@5."
Citazioni
"Coarse+fine (proposed) improved 9% and 12% in MRR and nDCG@5 compared to fine-tuned."
"Fine-tuning improved effectiveness with prior coarse-tuning."