information retrieval

This survey addresses the integration of information retrieval (IR) and natural language processing (NLP), focusing on systematic comparison across sparse, dense, and hybrid retrieval paradigms. Methodologically, it unifies classical and modern frameworks—including Lucene, Anserini, Pyserini—as well as foundational models (vector space, probabilistic, inference networks) and state-of-the-art pre-trained Transformers (e.g., BERT). Its primary contribution is the first comprehensive, cross-paradigm empirical evaluation of sparse, dense, and hybrid retrieval, extended to emerging application domains: cross-lingual IR, argument mining, privacy-preserving retrieval, and hate speech detection. The study clarifies co-evolutionary trajectories between IR and NLP, identifies core challenges in accuracy, scalability, and ethical robustness, and proposes future research directions toward trustworthy, fair, and efficient retrieval systems.

This work exposes a critical vulnerability in deep learning–driven dense embedding retrieval systems under malicious SEO attacks: adversaries can inject minimal adversarial text to hijack retrieval rankings for sensitive queries (e.g., celebrity names), causing malicious content to be prioritized. To address this, we propose GASLITE—the first black-box, gradient-driven adversarial text generation method that requires no access to training corpora, model parameters, or gradients. Leveraging geometric modeling of the embedding space, GASLITE enables mathematically verifiable rank manipulation. The generated adversarial passages preserve designated malicious semantics while achieving top-10 recall rates ≥90% for target concept queries. Evaluated across nine mainstream retrieval models, GASLITE improves attack success rate by ≥140% over baselines. Crucially, injecting adversarial text comprising ≤0.0001% of the corpus suffices to place malicious content within the top-10 results for 61–100% of unseen concept queries.

This paper systematically investigates the scalability of sparse text embeddings—based on SPLADE—in ultra-large-scale retrieval (MSMARCO-v2, 138 million documents). Addressing the efficiency-effectiveness trade-off of existing approximate search methods (e.g., Seismic hashing and graph-based indexing) for billion-scale sparse vectors, it provides the first empirical analysis of index construction cost, query latency, and recall. The contributions are threefold: (1) identifying critical bottlenecks in real-world industrial-scale sparse retrieval—namely, memory explosion in graph indexes and progressive accuracy degradation in Seismic hashing; (2) proposing an approximate nearest neighbor (ANN) optimization pathway specifically tailored to the structural properties of sparse embeddings; and (3) establishing the first end-to-end evaluation benchmark for billion-scale sparse embeddings, delivering empirically grounded insights and design principles for efficient, production-ready sparse retrieval systems.

This study investigates the impact of embedding dimensionality on performance in dense retrieval and its limitations as task complexity increases. Through systematic experiments across models of varying scales, the work presents the first empirical evidence that retrieval performance follows a power-law relationship with embedding dimensionality. Building on this observation, the authors propose predictable scaling laws based solely on dimensionality or jointly on model size. Using dense retrieval architectures, approximate nearest neighbor search, and large-scale comparative evaluations, they demonstrate that in task-aligned scenarios, performance improves with higher dimensionality—albeit with diminishing returns—whereas in misaligned tasks, excessive dimensions degrade performance. These findings offer both theoretical grounding and practical guidance for selecting optimal embedding dimensions in efficient retrieval systems.

To address the complexity and poor reproducibility of fine-tuning and inference engineering for Transformer-based models in information retrieval (IR), this paper introduces LightIR—a lightweight, open-source framework built on PyTorch Lightning. LightIR proposes a novel modular and unified end-to-end IR pipeline architecture, comprehensively supporting fine-tuning, indexing, retrieval, and re-ranking. It accommodates mainstream models (e.g., BERT, ColBERT) and enables distributed training and indexing. By abstracting generic interfaces and providing preconfigured templates, LightIR significantly lowers development barriers while enhancing experimental reproducibility and extensibility. Empirical evaluation on standard benchmarks—including MS MARCO and BEIR—demonstrates its effectiveness and efficiency. The framework is publicly released and has been widely adopted, filling a critical gap in accessible, high-performance IR experimentation frameworks.

This work proposes a novel approach to large-scale retrieval that circumvents the prohibitive cost of full reranking by constructing query and item embeddings derived from the outputs of a reranker. Specifically, it leverages relevance scores assigned by a heavyweight reranker over a set of support items to generate lightweight embeddings, thereby enabling the reranking model to directly guide embedding learning—a capability demonstrated here for the first time. Under mild conditions, the method is theoretically shown to approximate arbitrarily complex similarity functions. Through systematic investigation of support item selection strategies and integration with approximate nearest neighbor search, the approach significantly improves candidate set quality across multiple academic and industrial datasets while maintaining computational efficiency.

This study addresses the lack of systematic evaluation of hybrid search mechanisms that combine semantic retrieval with metadata filtering in existing vector databases. We propose a novel relevance metric, Global-Local Selectivity (GLS), construct MoReVec—the first benchmark dataset supporting filtered retrieval—and extend ANN-Benchmarks to enable unified evaluation of hybrid search performance. Through comprehensive experiments integrating diverse filtering strategies into FAISS, Milvus, and pgvector with IVFFlat and HNSW indexes, we demonstrate that engine-level algorithmic integration critically governs performance: Milvus achieves more stable recall via hybrid execution, pgvector’s optimizer often selects suboptimal query plans, and IVFFlat outperforms HNSW under low-selectivity queries. Our findings culminate in practical configuration guidelines that offer both theoretical insights and actionable recommendations for efficient hybrid search deployment.

This work proposes Embedding Inference Attacks (EIA), the first method capable of accurately identifying the backend embedding model in a strict black-box information retrieval setting where the adversary only receives unordered retrieved documents—without access to similarity scores or ranking information. By crafting discriminative queries, EIA effectively circumvents large language models’ refusal mechanisms for non-standard inputs and demonstrates feasibility even in realistic systems incorporating re-rankers or RAG architectures. Experimental results show that certain queries retain discriminative power despite re-ranking defenses, and the study further evaluates the limited efficacy of mitigation strategies such as similarity thresholds, thereby exposing critical model privacy risks in current RAG deployments.

This work proposes a semantic similarity computation method that integrates Word Mover’s Distance (WMD) with pretrained word embeddings such as GloVe to better model the semantic relationship between queries and documents in information retrieval. Traditional centroid-based word embedding approaches often fail to capture fine-grained semantic matches, particularly when handling synonymy and polysemy. By minimizing the transportation cost of aligning query and document terms in the embedding space, the proposed method achieves a more precise representation of semantic correspondence. Experimental results demonstrate that this approach significantly outperforms baseline models—including Doc2Vec and Latent Semantic Analysis (LSA)—on similarity ranking tasks, while maintaining domain independence and high retrieval accuracy, thereby confirming its effectiveness and generalizability in practical information retrieval scenarios.

This work proposes a novel approach to address the high computational overhead and latency inherent in multi-vector retrieval, which, despite improving recall, remains impractical for large-scale applications. By formulating multi-vector similarity search as a supervised learning problem, the method employs a single-hidden-layer neural network to map multi-vector representations into a compact single-vector embedding in a latent space. This transformation enables seamless integration with existing efficient approximate nearest neighbor search (ANNS) techniques. To the best of our knowledge, this is the first learnable framework that compresses multi-vector queries into single vectors while preserving high recall. Extensive experiments across diverse multi-vector text and vision models—including ColBERTv2—demonstrate the method’s effectiveness and generalizability, achieving up to an order-of-magnitude speedup in retrieval without sacrificing accuracy.