DisSim-FinBERT: Text Simplification for Core Message Extraction in Complex Financial Texts

This study proposes DisSim-FinBERT, a novel framework that integrates Discourse Simplification (DisSim) with Aspect-Based Sentiment Analysis (ABSA) to enhance sentiment prediction in complex financial texts. By simplifying intricate documents such as Federal Open Market Committee (FOMC) minutes, DisSim improves the precision of aspect identification, resulting in sentiment predictions that align more closely with economic events. The model preserves the original informational content and captures the inherent volatility of financial language, offering a more nuanced and accurate interpretation of long-form financial communications. This approach provides a practical tool for policymakers and analysts aiming to extract actionable insights from central bank narratives and other detailed economic documents.

💡 Research Summary

The paper introduces DisSim‑FinBERT, a novel framework that couples Discourse Simplification (DisSim) with Aspect‑Based Sentiment Analysis (ABSA) to improve sentiment prediction and aspect extraction in complex financial texts such as Federal Open Market Committee (FOMC) minutes. The authors begin by highlighting the pivotal role of central‑bank communication in shaping market expectations and monetary policy effectiveness, noting that existing models like BERT and its financial variant FinBERT struggle with the dense, multi‑aspect nature of these documents.

A comprehensive review of related work covers domain‑specific datasets, key‑phrase extraction, recent ABSA advances, and prior text‑simplification efforts, underscoring the need for a method that preserves informational content while reducing linguistic complexity.

The dataset consists of 32,034 sentences drawn from FOMC minutes spanning 2006‑2023. Of these, 1,030 sentences were manually annotated by three graduate researchers for six economic aspects (inflation, employment, growth, etc.) and three sentiment polarities (positive, neutral, negative). A majority‑vote scheme ensured label reliability; sentences lacking at least two agreeing annotators were excluded.

DisSim‑FinBERT’s architecture has two stages. The first stage, DisSim, employs a handcrafted rule set of 35 syntactic patterns derived from parse‑tree analyses. These rules recursively split complex sentences into minimal propositions, labeling clauses as either “core” (nucleus) or “satellite” (supporting context). The output is a hierarchical discourse tree with three levels: Level 0 (key statement), Level 1 (definitions/background), and Level 2 (evidence/details). The second stage fine‑tunes a pre‑trained FinBERT model on the DisSim‑processed sentences. Crucially, only Level 0 sentences receive sentiment labels, which mitigates cosine‑similarity‑driven misclassifications that occur when FinBERT treats nearby sentences as interchangeable. The model simultaneously predicts sentiment scores for each identified aspect, delivering a nuanced, multi‑dimensional sentiment profile.

Experimental evaluation compares the baseline FinBERT (trained on raw minutes) with DisSim‑FinBERT. Using aspect‑level F1 and overall sentiment accuracy as metrics, DisSim‑FinBERT achieves an 8.7‑percentage‑point gain in aspect F1 and a 6.3‑point increase in sentiment accuracy. Error analysis reveals that the hierarchical simplification isolates the “key‑statement” sentiment, reducing ambiguity and improving interpretability for policymakers and analysts.

Key contributions include: (1) a rule‑based discourse simplification pipeline that structurally decomposes dense financial prose while preserving meaning, (2) integration of hierarchical discourse information with ABSA to prevent sentiment dilution across sentences, and (3) a modular approach that can be applied to any FinBERT‑compatible downstream task. Limitations are acknowledged: the handcrafted rule set is language‑ and domain‑specific, requiring adaptation for non‑English or non‑U.S. central‑bank texts, and the approach does not yet automate rule discovery.

Future work proposes (i) learning rule patterns automatically via meta‑learning, (ii) extending the framework to multimodal inputs (tables, charts) common in economic releases, and (iii) deploying the system in real‑time policy‑monitoring dashboards for rapid sentiment‑driven market analysis. In sum, DisSim‑FinBERT offers a practical, high‑precision tool for extracting actionable insights from long‑form financial communications, bridging the gap between complex central‑bank narratives and actionable market intelligence.