Sep 2026· Journal of Management in Engineering· Vol 42· 0 citations· 34 references
TL;DR
This study develops a large language model (LLM)-assisted text-mining framework for extracting key factors from large-scale unstructured documents and demonstrates that the approach not only corroborates established risk factors but also reveals overlooked causes, highlighting the potential of LLM-driven text mining to enhance large-scale risk knowledge extraction.
Abstract
Construction quality risks critically affect both users’ experience and safety. Learning from the vast quality data generated by construction projects offers valuable opportunities for knowledge discovery, yet extracting reliable risk factors from unstructured texts remains challenging. Existing approaches often rely solely on term frequency and generic lexicons, overlooking causal relations and domain-specific nuances. In addition, the existing risk checklists insufficiently aligned with the liability clauses of inherent defects insurance (IDI). To address these limitations, this study develops a large language model (LLM)-assisted text-mining framework for extracting key factors from large-scale unstructured documents. The system integrates causal relation extraction powered by the LLM for obtaining explicit pairs, a domain-specific lexicon and synonym set derived from regulatory documents and expanded via LLM for improved segmentation and normalization, and an entropy-augmented term weighting scheme [term frequency–information entropy (TF-H)] to enhance the robustness of risk factor identification. Applied to 36,998 judicial judgments on construction quality disputes from 2000 to 2022 in China, the framework generates liability clause–aligned risk factor lists for IDI. The results demonstrate that the approach not only corroborates established risk factors but also reveals overlooked causes, highlighting the potential of LLM-driven text mining to enhance large-scale risk knowledge extraction. Furthermore, the findings provide practical implications for various stakeholders, including insurers, technical inspection institutions and project managers, benefiting them regarding the core workflows of IDI schemes-specifically inspection, underwriting, claims, and so on.
This study presents a novel, manually annotated dataset of 121 business news articles related to five major steel companies, using the Cambridge Risk Taxonomy, and provides empirical evidence on the opportunities and limitations of LLMs for both analytical and narrative forms of automated risk assessment.
Laleh Davoodi, Filip Ginter, Sima Salimi et al.· SN Computer Science· 0 citations
Failure Modes and Effects Analysis (FMEA) is a systematic methodology employed to identify, prioritize, and mitigate potential and existing failure modes within a system. Despite its efficacy, the manual acquisition and processing of failure data from extensive datasets is a resource‐intensive endeavor. In the digital era, vast quantities of consumer‐generated data regarding product performance are available on e‐commerce platforms, social media, and specialized forums. However, the sheer volume and unstructured nature of this “Big Data” render manual interpretation and classification practically unfeasible. Consequently, sophisticated text mining techniques are required to extract actionable intelligence from these massive review corpora. While the literature extensively discusses quality improvement via data mining, research integrating advanced text mining within the FMEA framework remains scarce. This gap is addressed by proposing POStagging2vec, a novel text mining methodology designed to extract the most representative failure‐defining sentences without the need for exhaustive manual review. POStagging2vec enhances the analytical process by measuring semantic similarity between review sentences and cluster labels augmented with salient parts of speech, specifically adverbs, adjectives, verbs, and compounds. To validate the proposed framework, a case study was conducted on Amazon customer reviews for a robot vacuum cleaner. From an initial dataset comprising 820 reviews and 5906 sentences, the methodology efficiently isolated the 140 most relevant sentences, leading to the identification and evaluation of 20 critical failure modes. The results demonstrate that this integrated approach allows organizations to rapidly detect significant failures and expedite the implementation of corrective actions, thereby enhancing product reliability and consumer satisfaction via a more effective and successful R&D process for their products.
S. Ayber, Ahmet Aydın, Gökhan Göksel et al.· Quality and Reliability Engi...· 0 citations
As experienced workers retire across industrialized economies, organizations risk losing procedural expertise that often remains tacit, undocumented or scattered across unstructured documents. Translating this knowledge into structured, machine-readable representations is difficult to scale, labor-intensive, and prone to inconsistency when done manually. This paper addresses the automated construction of knowledge graphs from natural language procedural descriptions, developing a generic approach for transforming unstructured expert knowledge into structured knowledge graphs that support downstream retrieval and question-answering applications. Three text-to-knowledge-graph approaches were designed, implemented, and systematically evaluated. The first employed a large general-purpose language model (Qwen3-32B) with a single-stage zero-shot extraction prompt, the second applied the same strategy using a smaller base model (Llama2-13B), and the third combined supervised fine-tuning of the smaller model on synthetic extraction data with a decomposed extraction architecture targeting one to two ontology elements per phase. All approaches were evaluated across six procedural descriptions spanning multiple technical domains, with chunk size, model temperature, and ontology detail as configuration parameters. Results were assessed for intrinsic quality and extrinsic fitness for use, measured by question-answering accuracy in a Graph-RAG application. The results demonstrate that extraction strategy is a more decisive factor than model capacity. The fine-tuned model achieved a Question Answering (QA) pass rate of 55.3%, compared to 46.8% for the large model and 33% for the small base model, outperforming the general-purpose model on both intrinsic quality metrics and downstream performance. Average node degree, duplication rate, and ontology simplicity emerge as the strongest predictors of retrieval performance. Shorter ontologies consistently outperformed richer ones, suggesting that ontology design should be driven by the intended downstream application rather than semantic completeness. Over-extraction of procedural steps did not degrade performance but instead improved retrieval robustness by providing additional entry points for graph traversal. The findings offer practical guidance for designing scalable, locally deployable knowledge graph construction pipelines for procedural texts under computational and confidentiality constraints.
Erik Sörqvist, Kenneth Obinna, Clara Bersch et al.· European Conference on Knowl...· 0 citations
Unstructured text data such as crime reports and witness statements often contain essential connections between crime entities such as suspects, weapons, locations, and crime types. However, it can be difficult to extract and analyze these linkages because they are often inserted within unstructured narratives. This paper presents a hybrid framework that combines crime-specific entity recognition with fine-tuned large language models (LLMs) to extract relationships from crime data. The pipeline initially employs a BERT-based model to identify crime-related entities, followed by a Tiny LLAMA and LoRA-optimized LLM to discern and structure relationships as subject–predicate–object triplets which helps in criminal analytics. Experiments conducted on selected crime dataset showed significant improvements in 85.8 % accuracy with F1-score, recall, and precision over baseline models, especially when handling complicated relationship types. The results show that transformer-based NER combined with targeted LLM fine-tuning offers an effective hybrid approach for extracting useful relations from unstructured crime texts, which is helpful to criminal justice analytics and law enforcement.
Sukhvinder Kaur Walia, S. Masih, U. Suman· International Journal of Lat...· 0 citations
This study develops an ontology-based event-intelligence framework for FinTech merger-and-acquisition evidence in U.S. Securities and Exchange Commission (SEC) structured disclosures. Five quarterly SEC Financial Statement Data Sets from 2025Q1 through 2026Q1 contain 32,254 filings, 7,335 registrants, and 18,312,494 numerical XBRL facts; validation adds a 1,800-filing full-text Form 8-K sample and 192 FDIC events. A deterministic ontology maps XBRL tag names, labels, and documentation to acquisition, disposition, valuation, integration, risk, and payment concepts. The method is therefore LLM-inspired semantic enrichment rather than direct LLM extraction. Logistic regression, decision tree, and random forest classifiers are evaluated in three expanding forward-quarter tests with training-only preprocessing and threshold selection. A proximal protocol retains semantically close predictors, whereas a strict protocol excludes label-generating variables and deterministic descendants. Across the rolling tests, proximal logistic-regression M&A detection attains mean F1 = 0.981941, ROC-AUC = 0.999517, and average precision = 0.998315; strict performance falls to F1 = 0.735640, ROC-AUC = 0.930870, and average precision = 0.824107. FinTech M&A F1 declines from 0.921198 to 0.449503. Strict random forests yield F1 = 0.798129 for integration risk and 0.753987 for valuation signals. In independent full text, strict main-text-plus-exhibit F1 is 0.297482; among 24 automatically linked FDIC events in rolling test quarters, 9 are detected. Near-perfect scores therefore describe ontology reconstruction, not transaction-level accuracy.
Guanzheng Zhao, Di Zhang, Sisi Meng· Journal of Technology Inform...· 0 citations
This research focuses on enterprise profiling in scenarios where large volumes of diverse texts—such as registration records, annual reports, news articles, and bidding notices—are continuously generated. Instead of relying solely on a single data representation or classification model, we developed a comprehensive natural language processing (NLP) pipeline for extracting key information and identifying industries. The pipeline consists of several steps. First, we use a BERT-BiLSTM-CRF model to identify core e/nterprise entities. Then, we combine TF-IDF with BERT embeddings to create a hybrid feature scheme that captures both lexical cues and contextual semantics. To address the challenge of imbalanced industry labels, we apply SMOTE in the dense semantic space and pair it with Focal Loss to enhance learning for minority classes. Additionally, we introduce a Stacking strategy to integrate outputs from different models, making predictions more stable. Tests on a self-compiled dataset covering ten national economic sectors and about 50,000 enterprises show that our method achieves a macro-F1 score of 95.4%. It outperforms traditional machine learning baselines and single deep learning models, offering more reliable recognition for minority classes. These results suggest that our framework is well-suited for applications such as supply chain partner discovery, industrial mapping, and targeted investment promotion.
Xin-Yi Xu· Applied and Computational En...· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.
