Abstract
There is a wealth of information about financial systems that is embedded in document collections. In this paper, we focus on a specialized text extraction task for this domain. The objective is to extract mentions of names of financial institutions, or FI names, from financial prospectus documents, and to identify the corresponding real world entities, e.g., by matching against a corpus of such entities. The tasks are Named Entity Recognition (NER) and Entity Resolution (ER); both are well studied in the literature. Our contribution is to develop a rule-based approach that will exploit lists of FI names for both tasks; our solution is labeled Dict-based NER and Rank-based ER. Since the FI names are typically represented by a root, and a suffix that modifies the root, we use these lists of FI names to create specialized root and suffix dictionaries. To evaluate the effectiveness of our specialized solution for extracting FI names, we compare Dict-based NER with a general purpose rule-based NER solution, ORG NER. Our evaluation highlights the benefits and limitations of specialized versus general purpose approaches, and presents additional suggestions for tuning and customization for FI name extraction. To our knowledge, our proposed solutions, Dict-based NER and Rank-based ER, and the root and suffix dictionaries, are the first attempt to exploit specialized knowledge, i.e., lists of FI names, for rule-based NER and ER.
Introduction
The 2008 financial crisis underscored a critical transparency gap in the global financial system. Much of the data regarding financial supply chains remains "buried" within hundreds of pages of semi-structured text in SEC filings, such as residential mortgage-backed securities (resMBS) prospectuses. While these documents are public, extracting structured data from them is a significant challenge for standard Natural Language Processing (NLP) tools.
General-purpose Named Entity Recognition (NER) systems often struggle with financial institution (FI) names due to three primary factors:
- Complex Naming Conventions: FI names are often long, include alphanumeric characters, and follow rigid legal templates (e.g., "AAMES MORTGAGE INVESTMENT TRUST 2006-1").
- Formatting Irregularities: Names frequently span multiple lines or appear in headers and tables without surrounding sentence context, as illustrated in Figure 2.
- Entity Resolution (ER) Ambiguity: Variations like "Wells Fargo," "Wells Fargo Bank, N.A.," and "Wells Fargo Bank National Association" must be mapped to a single real-world entity.
Methodology
The proposed framework, built on the System T declarative information extraction platform, utilizes a two-stage pipeline: Dict-based NER followed by Rank-based ER.
Dict-based NER
The core innovation is the "Root-Suffix" heuristic. Rather than treating an institution name as an atomic string, the system decomposes it into a Root (the unique identifier, e.g., "Wells Fargo") and a Suffix (the functional descriptor, e.g., "Bank" or "N.A."). By automatically generating extensive dictionaries—over 350,000 roots and 26,000 suffixes—from existing financial lists (SEC, ORG names, etc.), the system can "reconstruct" names. This approach is significantly more robust to line breaks and non-standard formatting than traditional sentence-based NER.
Rank-based ER
To resolve extracted mentions against a master list (such as the ABSNet corpus), a custom scoring function inspired by TF-IDF was developed. This function incorporates a weight decay mechanism that prioritizes the initial tokens (the root) and applies a "bonus" for substring matches. This ensures that "Wells Fargo" correctly resolves to "Wells Fargo Bank" with higher confidence than "Bank of America."
Evaluation and Results
The study utilized a corpus of over 5,000 resMBS prospectuses filed between 2000 and 2008. Performance was evaluated against a general-purpose "ORG NER" baseline.
The specialized Dict-based NER demonstrated superior performance:
- Precision: In "Both" (Header + Summary) evaluations, the specialized system achieved 94.15% precision, vastly outperforming the general system's 68.38%.
- Recall: Both systems maintained high recall (~96% for specialized vs. ~85% for general), but the specialized system produced significantly fewer fragmented or partial extractions.
- ER Performance: The Rank-based ER system achieved 99.95% precision at its optimal threshold, validating the effectiveness of the weighted scoring function in handling naming variations.
Conclusion
This research demonstrates that for specialized domains like finance, rule-based systems leveraging domain-specific knowledge can outperform general-purpose machine learning models. By moving away from sentence-boundary clues and focusing on the functional components of names, the system achieves high reliability in extracting data from complex document layouts. Future work will explore the use of regular expressions for serialized names and expanding the methodology to other asset classes such as auto or student loans.
Summary updated