Pseudo-Relevance Feedback (PRF) is a key technique in information retrieval (IR). Traditional implementations rely on statistical information, such as term frequency, for precise matching and relevance assessment. However, these methods struggle to fully capture the deep semantic integrity of query terms, especially in handling polysemy, high semantic relevance, and long-document comprehension. To address these challenges, this paper innovatively proposes a large language model-assisted PRF probabilistic model. The model first employs a precise matching algorithm to evaluate and determine the term-level weights, and then uses a large language model to encode the contextual relationships within the query and feedback documents, thereby accurately acquiring the global semantic weights of terms relevant to the query at the document level. By adjusting a balancing factor to allocate weights between these two components, the model comprehensively selects expanded terms for constructing a new query representation and executing query expansion (QE). This model not only facilitates approximate matching through the integration of global semantic features of documents but also effectively combines with the precise matching information of traditional PRF models, enabling a comprehensive and accurate optimization of queries from a broader perspective. To validate effectiveness, extensive empirical analyses on five TREC datasets assess performance across key metrics such as MAP, P@10, NDCG, and MRR. Experimental results show significant improvements over baseline models. Comparative analyses and case studies confirm that the expanded terms maintain high semantic relevance and consistency with the original query while preserving diversity and effectively capturing global document semantics, establishing an efficient QE mechanism.

In the field of information Retrieval (IR), Pseudo-relevance feedback (PRF) and Query Expansion (QE) techniques have garnered significant attention for their efficacy in enhancing retrieval effectiveness. However, traditional PRF approaches predominantly concentrate solely on pseudo-relevant documents identified during the initial retrieval stage, neglecting the rich semantic information embedded within non-pseudo-relevant documents. This paper introduces an innovative PRF model that integrates soft negative samples and contrastive learning to address this limitation, aiming for a more comprehensive capture and representation of semantics. First, we employ the BM25 algorithm as the baseline retrieval mechanism to accurately pinpoint pseudo-relevant documents from the first stage retrieval and assign weights to their terms. Second, a contrastive learning strategy is introduced to distill semantic features from all documents globally, further refining the semantic weights of terms. To mitigate the risk of information loss associated with soft negative samples, we ingeniously leverage the statistical properties of kernel function to precisely gauge the co-occurrence frequencies between terms, ensuring the preservation of core information and thus obtaining kernel function term co-occurrence weights. Third, we select semantically related terms highly relevant to the query for creating an optimized query by balancing these three weight distributions. Extensive empirical analyses conducted on several TREC datasets demonstrate the practical feasibility of our proposed model. It outperforms baseline models and state-of-the-art technologies on core evaluation metrics such as MAP, P@10, NDCG, and MRR. Deeper comparative experiments and case studies reveal that the expansion terms generated by our model exhibits a deeper level of semantic coherence with the original query, underscoring the dual advantages of the model in both theory and practice. In summary, the model presented herein not only opens a new path at the technical level, but also provides a more accurate and efficient solution for real-world applications in IR.

Conversational retrieval leverages multi-turn conversations to meet users’ information needs, and accurately understanding the new intent has become a significant challenge in this field. Recently, the language comprehension and reasoning capabilities of large language models (LLMs) offer a viable solution to these challenges. In this article, we propose a new Dual-Query Generation and Joint-Encoding method by utilizing LLM for Conversational Information Seeking, abbreviated as DQ-CIS. Specifically, we propose a dual-query generation approach that leverages both open source and closed source LLMs to generate two complementary queries: a full-rewrite query that preserves the context semantics of the conversation and a condensed-rewrite query that emphasizes the core intent of the current query. Additionally, to better express the semantic information of the query, we propose a dual-query joint-encoding method, which enhances the thematic expression of query vectors by treating the dual-query as semantic complementary. A query coverage fine-tuned semantic matching method is also introduced to improve result relevance and ranking by fine-tuning the original retrieval scores by ColBERT. We conducted a number of experiments on seven publicly available conversational retrieval datasets. The results show that compared with other models, DQ-CIS has strong competitiveness in both retrieval efficiency and retrieval results.

Abstract Recently, large pretrained language models (PLMs) have led a revolution in the information retrieval community. In most PLMs‐based retrieval frameworks, the ranking performance broadly depends on the model structure and the semantic complexity of the input text. Sequence‐to‐sequence generative models for question answering or text generation have proven to be competitive, so we wonder whether these models can improve ranking effectiveness by enhancing input semantics. This article introduces SE‐BERT, a semantically enhanced bidirectional encoder representation from transformers (BERT) based ranking framework that captures more semantic information by modifying the input text. SE‐BERT utilizes a pretrained generative language model to summarize both sides of the candidate passage and concatenate them into a new input sequence, allowing BERT to acquire more semantic information within the constraints of the input sequence's length. Experimental results from two Text Retrieval Conference datasets demonstrate that our approach's effectiveness increasing as the length of the input text increases.