Jul 2026· 2026 International Conference on Intelligent and Sustainable AI Systems (ICOSAAS)· pp. 1082-1086· 0 citations· 15 references
Abstract
Large Language Models are deployed in financial applications such as research synthesis and risk analysis, yet their effectiveness is constrained by the limitations of conventional retrieval methods. Existing approaches rely primarily on semantic similarity or token-level matching, which fails in structured domains like finance where relevance depends on precise alignment across entity, temporal and document-type dimensions. This paper proposes the Financial Knowledge Integration Framework (FKIF), a hybrid metadata-aware retrieval system that integrates dense semantic similarity, sparse lexical matching, and structured metadata signals into a unified ranking function. Unlike conventional hybrid retrieval approaches, FKIF treats metadata as a first-class relevance signal rather than an auxiliary feature. Tested on a held-out set of 22 queries drawn from a corpus of 57 SEC filings, FKIF achieves an MRR@10 of 0.98 against TF-IDF’s 0.35, BM25’s 0.17, and dense retrieval’s 0.13. The results demonstrate that metadata-aware retrieval significantly enhances retrieval accuracy and provides a foundation for reliable financial Retrieval-Augmented Generation (RAG) systems.
Information retrieval has changed dramatically over the past decades. Early systems relied on simple keyword matching, but modern search engines must understand meaning, context, and user intent. This paper examines three major families of retrieval models that have shaped this evolution: vector space models, probabilistic retrieval, and neural retrieval. Vector space models represent documents and queries as weighted term vectors and rank them by similarity, providing a simple yet effective way to handle partial matches. Probabilistic models, such as BM25, treat relevance as a probability and rank documents according to how likely they are to satisfy a query, offering a stronger theoretical foundation for ranking. Neural retrieval goes further by learning dense semantic representations that can capture meaning beyond exact word overlap, enabling more accurate matching and reranking. We review key works including Salton et al.’s foundational vector space model, Robertson and Zaragoza’s probabilistic relevance framework, and recent neural approaches such as Dense Passage Retrieval and large language model-based retrieval surveys. The discussion shows that modern search systems rarely rely on a single model. Instead, they combine fast lexical retrieval with powerful neural rerankers to balance speed and accuracy. This hybrid approach reflects the current state of the field and points toward future research directions.
Keywords: information retrieval; BM25; vector space model; probabilistic retrieval; neural retrieval; reranking
Prapitha Gopi K· International Journal of Tec...· 0 citations
Experiments on three real-world datasets demonstrate that AnnoIndex consistently outperforms state-of-the-art baselines, achieving the highest average F1 score while maintaining robust performance on complex multi-hop join and progressive reasoning queries.
Approaches to subject indexing need to consider a trade-off between operational efficiency and the maintenance of indexing quality. Given that existing automated approaches struggle with domain adaptability, this study proposes the recall–rank–rerank (R3) framework—a two-stage hybrid intelligence method for automated subject indexing. By simulating human indexing cognition through case-based reasoning and human–machine collaboration trained on the TIBKAT data set, the results demonstrate that R3 significantly improves efficiency and semantic precision within the standard test collection evaluation framework in information retrieval research. R3 operates without model fine-tuning: its first stage uses cross-lingual semantic embedding to recall and rank candidate subjects from similar documents, while the second employs a large language model (LLM) as a simulated expert for refinement—calibrating semantics, enriching implicit concepts, and reranking relevance. Results show R3 outperforms supervised fine-tuning and semantic recall (R2) methods. On the validation set, it lifts mean average precision (MAP) from 41.19% to 45.24% (a 4% gain) and increases top-5 precision (P@5) by 2.49%. In the LLMs4Subjects task, R3 achieves top recall rates—65.68% for core subjects and 58.56% for all subjects—leading the SemEval’25 shared evaluation. With its lightweight design and strong performance, R3 offers adaptability and scalability, delivering a possible solution for large-scale indexing in practice. Fusing semantic embedding with LLM reasoning, the proposed framework can boost accuracy and provide new paths for hybrid intelligence in artificial intelligence–assisted subject indexing. Implications for the design and evaluation of automated subject indexing systems are discussed.
T. Xia, Xin Yang, Wenjing Wu et al.· Journal of information scien...· 0 citations
Analyzing financial documents such as 10-K filings, tabular disclosures, and macroeconomic reports demands expert reasoning and extensive time. However, existing Retrieval-Augmented Generation systems often struggle to process hybrid text-table structures or the massive scale of financial documents. To address these challenges, we propose Hierarchical Reranker, a RAG framework designed to improve retrieval performance and generative reliability across large-scale financial datasets. The system integrates three key innovations: Pre-Retrieval Optimization, enhancing query clarity and search efficiency through normalization, keyword expansion, and table transformation; Hierarchical Reranker Architecture, improving retrieval precision through a two-stage ranking mechanism; and Long-Context Management, preserving reasoning accuracy through adaptive input partitioning and fusion under extensive contexts. Across multiple benchmarks, including FinQA, FinanceBench, and ConvFinQA, the proposed system achieved an NDCG@20 score of 0.7918 and demonstrated superior factual consistency. Its robustness was further validated by achieving second place in the ACM-ICAIF'24 FinanceRAG Challenge. This work presents a deployable, domain-optimized RAG pipeline that enhances both the accuracy and scalability of financial reasoning, paving the way for automated audit reporting and quantitative investment analysis. The source code will be made publicly available on GitHub upon acceptance.
The proposed REKALM, a comprehensive integration framework for enhancing LLM-based recommenders through knowledge integration, demonstrates that augmenting LLMs with lexicalized, domain-specific knowledge is an effective system-level strategy for advancing the next generation of recommender systems.
Alessandro Petruzzelli, C. Musto, Marco De Gemmis et al.· ACM Transactions on Informat...· 0 citations
H+ Embedding is introduced, a unified multi-granularity retriever that predicts variable-length phrase partitions, preserves uncovered tokens as singletons, and applies importance-guided unit selection with weighted MaxSim interaction.
Shusen Zhang, Junyi Hu, Ye Feng et al.· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.
