Technical Summary: Structured Data Extraction from Real Estate Documents using Clustering, Classification, and Large Language Models

1. Problem Statement

Real estate listings in the United Kingdom expose basic structured metadata (e.g., price, bedroom count) via APIs. However, the most critical property-level information—such as legal status, structural conditions, utility supplies, and heating systems—resides in attached questionnaire documents (PDFs). These documents are currently ignored by automated systems due to three primary challenges:

• Heterogeneity: Documents vary significantly in format, including digitally generated text, scanned physical forms, and complex layouts with checkbox annotations.
• Unstructured Nature: Information is often captured in free-form text with inconsistent terminology (e.g., "gas central heating" vs. "GCH" vs. "mains gas"), rendering traditional rule-based extractors (regex, template matching) ineffective and non-scalable.
• Lack of Processing: These documents are not stored in databases, accessible via APIs, or indexed by search engines, creating a "rich data gap" in the real estate pipeline.

2. Methodology

The authors propose an end-to-end pipeline applied to 3,965 questionnaire documents acquired from the Aberdeen Solicitors Property Centre (ASPC) platform. The pipeline consists of three sequential stages:

A. Data Acquisition

• Source: The ASPC platform, which lacks a public API.
• Technique: The authors reverse-engineered internal REST API endpoints using Chrome DevTools.
• Process: An authenticated session was established to access property listings. The system iteratively retrieved listing summaries, detailed attributes, and associated PDF URLs. A total of 3,965 unique property listings and their associated PDFs were downloaded.

B. Document Classification (Stage 1)

Before extraction, each PDF was categorized to determine the appropriate processing path:

1. Text Only: Digitally generated documents with selectable text.
2. Scanned: Image-rendered pages with no selectable text layer.
3. Special Char: Documents containing checkbox annotations or non-standard character encodings.

• Implementation: The classifier used pdfplumber to measure extractable text length (identifying scanned docs) and PyMuPDF to detect checkbox patterns. Ambiguous documents were further analyzed using pdf2image and Tesseract OCR solely for symbol detection.
• Outcome: 2,781 documents (70.1%) were classified as text_only and proceeded to extraction. The remaining 29.9% (scanned and special_char) were excluded in this study.

C. LLM-Based Structured Extraction (Stage 2)

• Model: DeepSeek R1 (specifically deepseek-r1-distill-llama-70b) accessed via the Groq inference API.
• Configuration: Temperature set to 0.3 to ensure output stability; top-p at 0.95; max 1,024 tokens.
• Prompt Engineering: The model was instructed to extract 35 predefined property attributes (covering legal status, structural condition, heating, utilities, etc.) and return them as a single, valid JSON object.
  • Constraints: Return only JSON (no preamble), represent all values as strings, use null for absent values, and include all 35 fields.
• Processing: Raw text from text_only PDFs was passed directly to the LLM without rule-based normalization. The output was parsed using regular expressions and inserted into a SQL database.
• Post-Processing: Duplicate listings were removed via exact address matching. Missing values were imputed as "No" (consistent with the questionnaire format), and features were normalized (StandardScaler/MinMaxScaler).

3. Key Contributions

1. Data Acquisition System: A reproducible pipeline using reverse-engineered REST APIs to acquire 3,965 heterogeneous questionnaire PDFs alongside structured metadata.
2. Document Classification Model: A routing mechanism that classifies documents into text_only, scanned, and special_char categories, directing them to appropriate processing pathways.
3. LLM Extraction Pipeline: A successful implementation using DeepSeek R1 to extract 35 structured attributes from 2,781 documents with a 100% success rate in generating valid JSON, eliminating the need for document-specific rule sets.
4. Downstream Validation: An empirical study demonstrating that the extracted features support similarity matching, market segmentation, and multi-criteria ranking.

4. Results and Validation

The pipeline processed 2,781 eligible documents, resulting in a final dataset of 2,766 unique property records enriched with 35 PDF-extracted features. The quality of the extracted data was validated through three downstream experiments:

• Feature Coherence (Similarity Matching):
  • Cosine similarity and Euclidean distance produced a Jaccard consistency score of 0.82 on Top-5 recommendations.
  • This high agreement indicates that the extracted feature vectors capture genuine relational structures between properties, rather than noise or hallucinations.
• Semantic Quality (Clustering):
  • K-Means clustering (k=2) yielded a Silhouette Score of 0.2088.
  • The resulting clusters were interpretable: Cluster 0 represented "Affordable/smaller properties," and Cluster 1 represented "Premium/larger properties." This suggests the extracted features reflect real-world market distinctions.
• Feature Richness (Multi-Criteria Decision Making):
  • Two ranking methods (Weighted Scoring and TOPSIS) applied to the same features produced low overlap (Jaccard ≈ 0.18) in their Top-5 lists.
  • The authors interpret this divergence as evidence that the dataset contains sufficient signal and variation to support nuanced, objective-dependent ranking, rather than forcing convergence on outliers.

5. Significance and Claims

The paper claims to demonstrate that LLM-based text extraction is feasible and reliable at scale for heterogeneous real estate documents.

• Feasibility: The system successfully processed thousands of documents without manual rule creation, handling terminological variations that defeat traditional parsers.
• Data Quality: The downstream validation provides empirical evidence that the extracted data is semantically meaningful and suitable for machine learning tasks (clustering, ranking, matching).
• Scalability: The pipeline successfully handled a live dataset of nearly 4,000 documents, proving that LLMs can be integrated into production pipelines for document intelligence.

6. Limitations and Future Work

The authors modestly acknowledge specific limitations:

• Selection Bias: Approximately 29.9% of documents (scanned and special_char) were excluded, limiting the dataset's representativeness.
• Lack of Ablation Study: No formal comparison was conducted between models trained on API-only data versus the PDF-enriched data to quantify the additive value of the extraction.
• Quantitative Validation Only: The study relies on quantitative metrics; no user study was conducted to evaluate the perceptual quality of the data.

Future Directions identified by the authors include:

• Integrating OCR preprocessing (noise reduction, deskewing) to handle scanned documents.
• Conducting formal ablation studies to measure the specific value added by PDF features.
• Adapting the domain-agnostic architecture to other sectors (legal, medical, insurance) and other real estate markets (commercial, rental).