A System for Name and Address Parsing with Large Language Models

Technical Summary: A System for Name and Address Parsing with Large Language Models

Problem Statement
The reliable transformation of unstructured person and address text into structured data remains a critical bottleneck in large-scale information systems. While traditional rule-based and probabilistic methods perform well on clean inputs, they often fail when confronted with noisy, multilingual, or inconsistently formatted data common in real-world scenarios (e.g., missing delimiters, typographical errors, and mixed languages). Conversely, while neural models and Large Language Models (LLMs) offer improved robustness, they frequently lack deterministic control, reproducibility, and strict adherence to output schemas, leading to structural or logical errors in generated data. Existing literature often relies on domain-specific fine-tuning, which limits adaptability across different countries and formats, or lacks post-generation validation layers to ensure cross-field consistency.

Methodology
The authors propose a prompt-driven, validation-centered framework designed to convert free-text records into a consistent 17-field schema without requiring model fine-tuning. The system operates as a four-stage pipeline using the Claude 4.0 Sonnet model via AWS Bedrock, configured with fixed decoding parameters to ensure reproducibility:

1. Input Normalization: Raw text (from files or free text) undergoes preprocessing to reduce variability. This includes Unicode canonicalization (NFKC), stripping zero-width characters, collapsing whitespace, and expanding abbreviations (e.g., "Ave" to "Avenue"). Directional tokens are mapped to a canonical set, and ordinal suffixes are normalized.
2. Prompt Assembly: A compact, template-based prompt guides the LLM. It includes a specific role definition, a strict 17-field schema order, explicit rules (e.g., "no invented data," "use empty strings for unknowns"), and few-shot examples covering edge cases (e.g., hyphenated numbers, pre/post directionals). The prompt is designed to keep the model in "extractor mode" rather than generative mode.
3. Constrained Inference: The model processes records in fixed batches of 16. Decoding parameters are strictly controlled (Temperature: 0.30, Top-p: 0.90) to favor consistency over creativity. Stop sequences ensure output terminates exactly at record boundaries. If a structural check fails (e.g., incorrect field count), a single retry with tighter constraints is attempted.
4. Strict Validation and Export: A deterministic validation layer operates post-generation. It enforces:
   • Schema Checks: Verifying exact field counts and data types.
   • Canonicalization: Mapping values to admissible sets (e.g., USPS state codes, long-form street types).
   • Cross-Field Rules: Enforcing real-world constraints, such as ZIP code and state compatibility.
     Records failing validation are flagged with specific reasons for assisted review. A lightweight confidence score is generated based on rule passes and token rarity.

Key Contributions

• Fine-Tuning-Free Pipeline: The system achieves high parsing accuracy without the need for large annotated corpora or domain-specific model retraining, relying instead on structured prompting.
• Deterministic Validation Layer: By integrating a strict rule-based validation step after LLM generation, the system combines the flexibility of generative models with the reliability of rule-based approaches, ensuring schema adherence and logical consistency.
• Reproducibility and Transparency: The study utilizes fixed experimental parameters (batch size, decoding settings, prompt templates) across all runs, addressing the lack of reproducibility often found in LLM-based extraction research.
• Robustness to Noise: The framework is specifically designed to handle heterogeneous, noisy inputs, including multilingual data and incomplete records, rather than relying solely on benchmark-specific clean data.

Results
Evaluated on a corpus of 1,500 unique person–address records (including U.S., Puerto Rican, international, and synthetic noisy data), the system demonstrated:

• Overall Accuracy: 99.8% exact-row accuracy, representing a 0.46 percentage point improvement over a pattern-based baseline (99.34%) and a ~23% reduction in error rate.
• Field-Level Performance: Core fields (names, street numbers, states) achieved near-ceiling accuracy (99.7%–99.9%). Even traditionally variable fields like directionals and unit markers remained above 99%.
• Automation: The system automatically processed 94% of previously exception-prone records without manual intervention.
• Confidence Calibration: Predicted confidence scores aligned well with empirical correctness. Records with confidence above 90% showed 99.9% accuracy, allowing downstream processes to prioritize only the small subset of low-confidence records (4%) for verification.
• Error Analysis: The residual 0.2% error rate was primarily attributed to irregular formatting disrupting pattern recognition (45%) and record separation failures in unstructured text (35%).

Significance and Claims
The paper claims that combining structured prompting with deterministic validation provides a robust, interpretable, and scalable solution for structured information extraction. The authors assert that this approach offers a practical alternative to training-heavy or domain-specific models, making LLM-based extraction systems more viable for production environments where precision, consistency, and reproducibility are essential. The work demonstrates that high reliability can be achieved without retraining, offering a cost-efficient path to deployable structured data extraction at scale.