Abstract
The financial sector generates vast quantities of unstructured data, ranging from regulatory filings to earnings call transcripts. Processing this data requires high precision and the ability to handle thousands of specialized labels. This paper introduces two novel frameworks, FLAN-FinXC and FLAN-FinBPS, which leverage instruction-tuned Large Language Models (LLMs) and Low-Rank Adaptation (LoRA) to automate complex financial tasks. By shifting from traditional discriminative classification to a generative approach, FLAN-FinXC achieves state-of-the-art performance in extreme numeral labeling, particularly for unseen tags. Simultaneously, FLAN-FinBPS addresses the challenges of long-form summarization through a two-stage extractive-abstractive pipeline. Our results demonstrate significant improvements in both semantic accuracy and factual consistency across major financial benchmarks.
1. Introduction
Financial analysts and regulators are currently overwhelmed by the volume of periodic filings (such as SEC 10-Q and 10-K reports) and Earnings Call Transcripts (ECTs). Two primary bottlenecks hinder the automation of these workflows:
- Extreme XBRL Tagging: The eXtendable Business Reporting Language (XBRL) is the standard for tagging numerical values in financial reports. However, taxonomies are massive; for instance, the FNXL dataset contains 2,794 unique labels. Traditional discriminative models struggle with this "long tail," often failing to identify labels that were infrequent or absent during training.
- Long-Form Summarization: Earnings calls are lengthy, unstructured teleconferences with high compression ratios—meaning they contain significant "fluff" surrounding a few critical facts. These transcripts often exceed the token limits of standard LLMs, making direct summarization difficult.
To address these issues, we propose a shift toward generative instruction tuning, utilizing the FLAN-T5 architecture and parameter-efficient fine-tuning (PEFT) via LoRA.
2. Methodology
2.1 FLAN-FinXC: Generative Numeral Labeling
FLAN-FinXC treats XBRL tagging as a generative task rather than a multi-class classification problem. This allows the model to leverage the semantic meaning of the tags themselves.
- The Pipeline: The model is provided with an instruction, a specific financial statement context, and a question regarding a numerical value. Instead of outputting a class ID, the model generates the natural language documentation associated with the correct XBRL tag.
- Tag Matcher: To ensure the output aligns with official taxonomies, a separate "Tag Matcher" module maps the LLM's generated text to the most semantically similar official XBRL label. This allows for "zero-shot" labeling, where the model can correctly identify a concept it never encountered during training by understanding its documentation.
2.2 FLAN-FinBPS: Bullet-Point Summarization
FLAN-FinBPS is a two-stage framework designed to handle the length and complexity of earnings calls:
- Stage 1 (Extractive): An unsupervised, question-based module identifies core financial pillars (e.g., revenue, earnings per share, equity) and extracts the most relevant sentences from the transcript.
- Stage 2 (Abstractive): An instruction-tuned module takes the extracted context and transforms it into concise, factually consistent bullet points. This prevents the model from losing focus over thousands of words of transcript text.
3. Experimental Setup and Datasets
We evaluated our frameworks using several specialized financial benchmarks:
- FNXL & FiNER: Datasets for extreme numeral labeling. FNXL is particularly rigorous, featuring nearly 2,800 unique labels.
- ECTSum: A dataset containing over 2,400 pairs of long earnings call transcripts and expert-written bulleted summaries.
All models were fine-tuned using LoRA, which significantly reduces the number of trainable parameters, making the tuning process computationally efficient while maintaining the performance of the underlying FLAN-T5 model.
4. Results and Discussion
4.1 Performance in Extreme Labeling
FLAN-FinXC demonstrated a remarkable ability to handle the "long tail" of financial labels. It achieved a 58.89% Macro-F1 score on XBRL tags that were completely unseen during the training phase. This represents a 41% improvement in Macro-F1 for rare tags compared to the previous state-of-the-art model, AttentionXML. By generating descriptions rather than IDs, the model bridges the gap between raw numbers and complex accounting taxonomies.
4.2 Summarization and Factuality
In the summarization task, FLAN-FinBPS outperformed the strongest baselines across multiple metrics:
- ROUGE Scores: Increased by 14.88%.
- BERTScore: Increased by 16.36%.
- Factuality: Measured by SummaC_CONV, factual consistency improved by 2.70%. In the financial domain, where a single hallucinated digit can lead to incorrect valuations, this improvement in consistency is critical.
5. Conclusion and Future Work
This research demonstrates that generative instruction tuning is superior to traditional discriminative methods for complex financial tasks. By focusing on semantic documentation and topic-driven extraction, our frameworks provide a robust solution for XBRL tagging and transcript summarization.
Future work will focus on expanding these frameworks to more recent LLMs, such as Mistral and Llama, and incorporating external financial knowledge bases to further reduce hallucinations. Additionally, we plan to explore the FinQA dataset to enhance the models' underlying numerical reasoning capabilities.
Summary updated