XBRLTagRec: Domain-Specific Fine-Tuning and Zero-Shot Re-Ranking with LLMs for Extreme Financial Numeral Labeling

1. Problem Statement

Publicly traded companies must disclose financial information using the eXtensible Business Reporting Language (XBRL), a standardized XML format. A critical step in this process is automatically assigning the correct XBRL tag to specific numerals found in financial text (e.g., matching "1.2 billion" to "Fair value of investment securities").

Key Challenges:

• Extreme Similarity: XBRL taxonomies contain thousands of tags with highly similar semantic definitions (e.g., distinguishing between "Fair value of investments" and "Fair value of investments in equity method"), making fine-grained differentiation difficult.
• Limitations of Existing Methods:
  • NER/Classification: Traditional Named Entity Recognition (NER) or extreme multi-label classification methods struggle with the vast tag space and fail to leverage the rich semantic structure of tag definitions.
  • Static Encoding: Many approaches rely on static tag encodings that do not fully utilize the descriptive text of XBRL tags.
  • Ranking Bottlenecks: Generative models can produce candidate tags but often lack robust mechanisms to rank and filter highly similar candidates to find the single optimal match.

2. Methodology: The XBRLTagRec Framework

The authors propose XBRLTagRec, an end-to-end framework that integrates three distinct modules to solve the tagging problem:

A. Tag Document Generation (Fine-Tuning)

• Model: Uses FLAN-T5-Large fine-tuned with LoRA (Low-Rank Adaptation) to ensure parameter efficiency.
• Input: A concatenation of an instruction prompt ($P$), the financial statement text ($D_i$), and a specific question targeting a numeral ($Q_x$).
• Output: Instead of predicting a raw tag ID, the model generates a semantic tag document (a natural language description of the tag). This approach leverages the richer semantic information contained in tag definitions rather than just labels.

B. Semantic Similarity Retrieval

• Encoder: Uses Sentence-T5-XXL to generate dense vector embeddings.
• Process: The generated tag document is encoded and compared against embeddings of all ground-truth XBRL tag documents in the taxonomy.
• Selection: The system retrieves the Top-10 most similar candidate tag documents based on cosine similarity. This narrows the search space from thousands to a manageable set of high-probability candidates.

C. LLM Zero-Shot Re-Ranking

• Mechanism: A large language model (ChatGPT-3.5) acts as a zero-shot re-ranker to resolve ambiguities among the Top-10 candidates.
• Iterative Strategy:
  1. Filtering: The Top-10 candidates are split into two groups (e.g., 5 documents each).
  2. Re-ranking: The LLM is prompted to rank these groups based on relevance to the generated tag document.
  3. Iteration: This process is repeated for multiple rounds (e.g., 1–8 iterations) with different random groupings to mitigate local biases.
  4. Voting: A frequency voting mechanism counts how often each candidate is ranked #1 across all iterations. The document with the highest frequency is selected as the final prediction.

3. Key Contributions

1. End-to-End Framework: Proposes XBRLTagRec, which bridges generative capabilities (LLM generation) with retrieval (semantic similarity) and iterative refinement (zero-shot re-ranking).
2. Semantic Document Generation: Shifts the paradigm from predicting discrete tags to generating semantic tag documents, allowing for finer-grained distinction between similar concepts.
3. Iterative Re-Ranking with Voting: Introduces a novel multi-round re-ranking strategy using frequency voting to stabilize predictions and reduce the impact of local ranking biases in LLMs.
4. Modular Architecture: The framework is designed with clear boundaries between generation, retrieval, and ranking modules, allowing for flexible swapping of generators, encoders, and re-rankers (demonstrated by testing with GPT-4, DeepSeek, and ERNIE).

4. Experimental Results

Experiments were conducted on the FNXL dataset (derived from SEC 10-K filings), which contains ~79k sentences and ~143k annotated numerals.

• Performance vs. Baselines: XBRLTagRec outperformed the state-of-the-art FLAN-FinXC and other baselines (FiNER, GalaXC, AttentionXML).
  • Hits@1: Improved by 2.64% (from 0.8354 to 0.8618).
  • Macro-F1: Improved by 4.34% (from 0.4616 to 0.5050).
  • Macro-Precision/Recall: Showed consistent gains of ~4.4%.
• Ablation Studies (RQ2):
  • Iterations: Performance peaked at 1–8 iterations; excessive iterations (beyond 8) led to diminishing returns or slight declines due to redundant information.
  • Group Size: A group size of 5 (splitting Top-10 into two groups of 5) yielded the best balance between performance and computational cost.
  • Ordering: Preserving the original cosine similarity order within groups (Order-Preserving) significantly outperformed random shuffling.
• Generalization (RQ3): The framework demonstrated robustness across different LLMs (ChatGPT-3.5, GPT-4, DeepSeek-V3, ERNIE-3.5), with GPT-4 achieving the highest absolute scores, confirming the model's plug-and-play adaptability.

5. Significance and Conclusion

• Practical Impact: The framework addresses a critical bottleneck in financial reporting automation, reducing the manual effort required for XBRL tagging and minimizing annotation errors caused by semantic ambiguity.
• Scalability: By combining efficient retrieval with powerful LLM reasoning, the system scales effectively to large taxonomies where traditional classifiers fail.
• Future Directions: While effective, the current reliance on iterative API calls (ChatGPT) incurs computational costs. Future work aims to explore efficient zero-shot methods and integrate external knowledge bases to further reduce latency and cost.

In summary, XBRLTagRec represents a significant advancement in financial NLP by effectively leveraging the semantic reasoning capabilities of LLMs to solve the "needle in a haystack" problem of matching financial numerals to highly similar standardized tags.