Abstract
Extracting precise information from financial documents is a critical yet challenging task due to the scarcity of labeled data and the complexity of domain-specific language. This paper presents DEXTER, a framework that combines deep learning with domain knowledge engineering to address Named Entity Recognition (NER) in low-resource settings. By utilizing a specialized Bi-LSTM-CRF architecture with domain-specific features and transfer learning, DEXTER achieves significant performance gains over standard models, reaching an F1 score of 87.48 on financial NER tasks.
Introduction
The financial sector relies heavily on the accurate extraction of entities—such as "Custodians" or "Asset Managers"—from lengthy, unstructured documents like fund prospectuses. Despite advancements in Natural Language Processing (NLP), the financial domain remains a "low-resource" environment. Data is often proprietary, jargon is dense, and the cost of expert annotation is prohibitive. Standard NER tools trained on general news corpora often fail to generalize to the specific linguistic structures of financial legalese.
Methodology
The DEXTER framework is built upon a Bi-directional Long Short-Term Memory (Bi-LSTM) network coupled with a Conditional Random Field (CRF) layer. To overcome data scarcity, the authors introduce two key innovations:
1. Domain-Specific Knowledge Engineering
Financial entities are frequently associated with specific "anchor" keywords. For instance, the phrase "Management Company" often signals the presence of an Asset Manager. DEXTER incorporates a dedicated Domain-Specific Bi-LSTM layer that encodes these expert-defined keywords as embeddings, injecting them directly into the neural architecture.
2. Layered Transfer Learning
Rather than training from scratch, the model is pre-trained on the large-scale CoNLL 2003 general-purpose NER dataset. The weights of the word and character embedding layers are then transferred to the financial model. This provides a foundational understanding of English syntax before the model is fine-tuned on specialized financial data.
Datasets and Benchmarks
The researchers evaluated DEXTER using five datasets:
- CoNLL 2003: Used for initial pre-training.
- Financial NER: A public dataset derived from SEC filings.
- New Low-Resource Benchmarks: Three manually annotated datasets from mutual fund prospectuses: Custodian (574 sentences), Asset Manager (672 sentences), and Leverage Ratio (140 sentences).
Results and Discussion
DEXTER consistently outperformed baseline models across all evaluated tasks. The inclusion of domain-specific features proved particularly effective in extremely low-resource scenarios. For the "Leverage Ratio" task, which had the smallest training set, the model achieved a peak F1 score of 95.65.
Ablation studies (shown above) demonstrate that the combination of word, character, and domain-specific features yields robust performance. Interestingly, the authors found that while transferring lower-level embeddings was beneficial, transferring the final projection layer from the general domain actually hindered performance, suggesting that the final classification stage must remain highly specialized to the target domain.
Conclusion
DEXTER demonstrates that integrating domain knowledge into modern deep learning architectures is a powerful strategy for low-resource NLP. While the current model relies on Bi-LSTMs, future work could explore the integration of Transformer-based models like BERT, as well as the incorporation of visual document features (e.g., layout and formatting) to further enhance extraction accuracy in complex financial documents.
This summary is based on the research paper "DEXTER - Data EXTraction & Entity Recognition for Low Resource Datasets" by researchers at Stride.AI.
Summary updated