A multi-dimensional semantic pseudo-relevance feedback framework for information retrieval.

Area of Science:

• Information Retrieval
• Natural Language Processing
• Machine Learning

Background:

• Pre-trained models are key for improving document ranking in information retrieval.
• Current methods often limit semantic information from pre-trained models to document re-ranking, neglecting query expansion.
• Pseudo-relevant documents contain crucial semantic information for effective query expansion.

Purpose of the Study:

• To propose a novel approach leveraging pre-trained models for multi-dimensional semantic information extraction from pseudo-relevant documents.
• To enhance query expansion by utilizing richer semantic insights.
• To improve overall information retrieval performance.

Main Methods:

• Initial retrieval using traditional sparse methods (e.g., BM25) for efficiency.
• Pre-trained models encode queries and document passages to extract sentence-level and passage-level semantic similarities.
• Combining term-level weights with extracted semantic weights to generate an improved query for a second retrieval round.

Main Results:

• Experiments on five TREC datasets and a medical dataset demonstrated significant improvements.
• Key metrics such as Mean Average Precision (MAP) and Precision at 10 (P@10) showed positive gains.
• The approach effectively utilized multi-dimensional semantic information for query expansion.

Conclusions:

• Leveraging multi-dimensional semantic information from pseudo-relevant documents significantly optimizes query expansion.
• The proposed method offers a more effective way to harness semantic information for enhanced retrieval performance.
• This study provides valuable insights into improving information retrieval systems through advanced semantic analysis.