Journals / Nicel Bilimler Dergisi / 2026 / Cilt: 8 Sayı: 1

Transforming Unstructured Corporate Documents into Queryable Knowledge Bases via Data Mining and Large Language Models

Yapılandırılmamış Kurumsal Dokümanların Veri Madenciliği ve Büyük Dil Modelleri ile Sorgulanabilir Bilgi Tabanlarına Dönüştürülmesi

Ali Sarıtaş(ESKISEHIR OSMANGAZI UNIVERSITY)0009-0007-2619-042X
İlknur Kaya(ESKISEHIR OSMANGAZI UNIVERSITY)0009-0007-5924-2265
Efnan Şora Günal(ESKISEHIR OSMANGAZI UNIVERSITY)0000-0001-6236-174X
Pages
100–117
Publication Date
2026-06-29
🏛️Bibliographic metadata and full text access for this article are provided via DergiPark.
DergiPark Record ↗
✨

AI Summary & Key Takeaways

AI Insight
⚡ TL;DR(Too Long; Didn't Read)

A comparative analysis based on concrete question-answering scenarios revealed that while well-structured data significantly enhances LLM accuracy, overall performance is heavily constrained by the models' context length limits and multi-row reasoning capabilities.

📌 3 Key Takeaways

🎯
Objective & Hypothesis:Unstructured corporate documents, such as invoices and inventory reports, pose significant challenges for automated data extraction and analysis.
📊
Key Findings & Data:Subsequently, the structured datasets were utilized as knowledge bases for five distinct large language models (LLMs), including GPT-5, Gemini 2.5, Grok 4, Qwen3-Max, and DeepSeek-V3.1 to evaluate their natural language query performance.
💡
Conclusion & Impact:A comparative analysis based on concrete question-answering scenarios revealed that while well-structured data significantly enhances LLM accuracy, overall performance is heavily constrained by the models' context length limits and multi-row reasoning capabilities.
🔬Design:Modeling & Simulation
📍Scope:Comprehensive Sample
⚙️Variables:Büyük Dil Modeli · Veri Madenciliği · Bilgi Çıkarımı

Abstract

Unstructured corporate documents, such as invoices and inventory reports, pose significant challenges for automated data extraction and analysis. This study proposes an end-to-end framework that transforms text-based PDF documents into highly structured CSV datasets using domain-specific regular expressions (regex) and data mining techniques, bypassing the need for optical character recognition (OCR). To validate the analytical consistency of the extracted data, descriptive statistical analyses, feature engineering, and customer segmentation were performed. Subsequently, the structured datasets were utilized as knowledge bases for five distinct large language models (LLMs), including GPT-5, Gemini 2.5, Grok 4, Qwen3-Max, and DeepSeek-V3.1 to evaluate their natural language query performance. A comparative analysis based on concrete question-answering scenarios revealed that while well-structured data significantly enhances LLM accuracy, overall performance is heavily constrained by the models' context length limits and multi-row reasoning capabilities. GPT-5 and Gemini 2.5 achieved the highest success rates. Ultimately, the findings demonstrate that coupling traditional data mining techniques with generative AI provides a highly reliable approach for transforming unstructured corporate documents into queryable decision-support systems.

Özet

Fatura ve stok raporları gibi yapılandırılmamış kurumsal dokümanlar, otomatik veri çıkarımı ve analizi açısından önemli zorluklar barındırmaktadır. Bu çalışma, metin tabanlı PDF dokümanlarını optik karakter tanıma (OCR) işlemine gerek kalmadan, alana özgü düzenli ifadeler (regex) ve veri madenciliği teknikleri kullanarak yüksek oranda yapılandırılmış CSV veri setlerine dönüştüren uçtan uca bir çerçeve önermektedir. Çıkarılan verilerin analitik tutarlılığını doğrulamak amacıyla tanımlayıcı istatistiksel analizler, öznitelik mühendisliği ve müşteri segmentasyonu işlemleri gerçekleştirilmiştir. Ardından, üretilen bu yapılandırılmış veri setleri, doğal dil sorgulama performanslarını değerlendirmek üzere GPT-5, Gemini 2.5, Grok 4, Qwen3-Max ve DeepSeek-V3.1 olmak üzere, beş farklı büyük dil modeli (LLM) için bilgi tabanı olarak kullanılmıştır. Somut soru-cevap senaryolarına dayalı karşılaştırmalı analizler, iyi yapılandırılmış verilerin LLM doğruluğunu önemli ölçüde artırdığını; ancak genel performansın modellerin bağlam uzunluğu sınırları ve çoklu satır ilişkilendirme yetenekleriyle güçlü bir şekilde kısıtlandığını göstermiştir. En yüksek başarı oranlarına GPT-5 ve Gemini 2.5 ulaşmıştır. Sonuç olarak elde edilen bulgular, geleneksel veri madenciliği tekniklerinin üretken yapay zekâ ile birleştirilmesinin, yapılandırılmamış kurumsal dokümanları sorgulanabilir karar destek sistemlerine dönüştürmek için son derece güvenilir bir yaklaşım sunduğunu kanıtlamaktadır.

🔖

Cite This Article

Sarıtaş, A., Kaya, İ., Günal, E. Ş. (2026). Transforming Unstructured Corporate Documents into Queryable Knowledge Bases via Data Mining and Large Language Models. Nicel Bilimler Dergisi, Cilt: 8 Sayı: 1, 100–117. https://doi.org/10.51541/nicel.1932954

📄 Full Text PDF Viewer

📄

PDF görüntüleyici hazırlanıyor...

References (13)

  1. Abdullah, M. H. A., Aziz, N., Abdulkadir, S. J., Alhussian, H. S. A. and Talpur, N. (2023), Systematic literature review of information extraction from textual data: recent methods, applications, trends, and challenges, IEEE Access, 11, 10535–10562. https://doi.org/10.1109/ACCESS.2023.3240898
  2. Bergling, O. (2025), A case study: How useful are RAG-LLM systems for enterprises?, Master’s thesis, University of Skövde, Skövde, Sweden.
  3. Cherguelaine, A. (2023), Company documents dataset, Kaggle, https://www.kaggle.com/datasets/ayoubcherguelaine/company-documents-dataset, Accessed: 25.06.2026.
  4. Gaikwad, S.V., Chaugule, A. and Patil, P. (2014), Text mining methods and techniques, International Journal of Computer Applications, 85(17), 42-45. https://doi.org/10.5120/14937-3507
  5. Lu, S., Su, Y., Zhang, X., Chai, J. and Yu, L. (2025), LLM-infused bi level semantic enhancement for corporate credit risk prediction, Information Processing & Management, 62(4), 104091. https://doi.org/10.1016/j.ipm.2025.104091
  6. Mohammed, H. H. (2019), Multi-label classification of text documents using deep learning, Master’s thesis, Cankaya University, Ankara, Türkiye.
  7. Rafidhul Haque, I. C. (2025), Implementation of retrieval-augmented generation (RAG) and large language models (LLM) for a document and tabular-based chatbot system, Journal of Electronics Technology Exploration, 3(1), 19–23. https://doi.org/10.52465/joetex.v3i1.588
  8. Scius-Bertrand, A., Jungo, M., Vögtlin, L., Spat, J. M. and Fischer, A. (2024), Zero-shot prompting and few-shot fine-tuning: Revisiting document image classification using large language models, Lecture Notes in Computer Science, 15319, Springer, Cham. https://doi.org/10.1007/978-3-031-78495-8_10
  9. Sharma, R. K. (2024), Power of design documents: building a feature-rich gen-ai chatbot with Python, OpenSearch, and LLMs, International Journal of Computer Applications, 186(55), 47-52. https://doi.org/10.5120/ijca2024924277
  10. Singh, A. P. (2025), Generative AI chatbot using CSV files, Authorea. https://doi.org/10.22541/au.176463848.81242143/v1
  11. Taşdelen, B. (2025), AI-powered project information systems: A chatbot model based on Horizon Europe projects, Master’s thesis, Ankara University, Ankara, Türkiye.
  12. Toprak, A. and Turan, M. (2023), Enhanced named entity recognition algorithm for financial document verification, Journal of Supercomputing, 79, 19431–19451. https://doi.org/10.1007/s11227-023-05371-4
  13. Yılmaz, Ş. (2020), A deep learning analysis on Turkish question classification task using word2vec representations, Doctorate thesis, Düzce University, Düzce, Türkiye.