Core Concepts
Coarse-tuning bridges pre-training and fine-tuning, improving effectiveness in ad-hoc document retrieval tasks.
Abstract
Introduction to Coarse-Tuning for Ad-hoc Document Retrieval using Pre-trained Language Models.
Proposal of Query-Document Pair Prediction (QDPP) for coarse-tuning.
Evaluation experiments showing significant improvements in MRR and nDCG@5.
Importance of learning query representations and query-document relations.
Experimental setup and results on various datasets.
Evaluation of query representation and query-document relations through prediction tasks.
Related work overview and conclusion with future research directions.
Stats
適切性を予測するQuery–Document Pair Prediction(QDPP)を提案。
4つのad-hocドキュメント検索データセットで、提案手法がMRRおよびnDCG@5を有意に向上させることを示す評価実験。
ORCASは1,900万のクリックされたクエリ–ドキュメントペアを含む。
Quotes
"Coarse-tuning helps to improve the effectiveness of the downstream IR tasks."
"Predicted tokens suggest that query representations and query-document relations were learned in coarse-tuning."