Scholarlink

<p>Recently, many approaches have addressed the problem of query expansion in information retrieval field, most of these approaches attempt to find best words to add them to the original query, without taking the appropriateness between these words. However, recent advances in deep learning and large language models have revolutionized query expansion approaches. State-of-the-art methods now leverage transformer architectures, semantic embeddings, and reinforcement learning to achieve superior performance. Despite these advances, current neural approaches often lack the global optimization capabilities of metaheuristic algorithms and fail to address the multi-objective nature of query expansion, which involves balancing precision, recall, diversity, and computational efficiency simultaneously. Consequently, we propose a hybrid approach that combines semantic understanding with metaheuristic optimization through a multi-stage architecture: (1) semantic candidate generation using pre-trained BERT models, (2) multi-objective fireworks optimization balancing precision, recall, diversity, and efficiency, (3) reinforcement learning fine-tuning for dynamic adaptation, and (4) attention-based query-document matching for enhanced relevance scoring. The proposed algorithm H-DFWA (Hybrid Deep Fireworks Algorithm) has been evaluated using seven benchmark datasets including CACM, TREC DL 2019/2020, MS MARCO, and BEIR. Experimental results demonstrate significant improvements: 75.7% enhancement over baseline on CACM (MAP: 0.724 vs. 0.412), 23% improvement in statistical significance across all datasets, and superior performance compared to state-of-the-art methods including BERT-QE, T5-QE, and HyDE.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Hybrid deep semantic query expansion using multi-objective fireworks-transformer optimization: a reinforcement learning approach

  • 期刊论文

  • 发表日期：

    2025年11月25日

查看全文
  • Fouad Bekkari,
  • Med Redouane Kafi,
  • Amine Khaldi

摘要

Recently, many approaches have addressed the problem of query expansion in information retrieval field, most of these approaches attempt to find best words to add them to the original query, without taking the appropriateness between these words. However, recent advances in deep learning and large language models have revolutionized query expansion approaches. State-of-the-art methods now leverage transformer architectures, semantic embeddings, and reinforcement learning to achieve superior performance. Despite these advances, current neural approaches often lack the global optimization capabilities of metaheuristic algorithms and fail to address the multi-objective nature of query expansion, which involves balancing precision, recall, diversity, and computational efficiency simultaneously. Consequently, we propose a hybrid approach that combines semantic understanding with metaheuristic optimization through a multi-stage architecture: (1) semantic candidate generation using pre-trained BERT models, (2) multi-objective fireworks optimization balancing precision, recall, diversity, and efficiency, (3) reinforcement learning fine-tuning for dynamic adaptation, and (4) attention-based query-document matching for enhanced relevance scoring. The proposed algorithm H-DFWA (Hybrid Deep Fireworks Algorithm) has been evaluated using seven benchmark datasets including CACM, TREC DL 2019/2020, MS MARCO, and BEIR. Experimental results demonstrate significant improvements: 75.7% enhancement over baseline on CACM (MAP: 0.724 vs. 0.412), 23% improvement in statistical significance across all datasets, and superior performance compared to state-of-the-art methods including BERT-QE, T5-QE, and HyDE.