Multi-Task Learning for Features Extraction in Financial Annual Reports

TL;DR

This work addresses the challenge of extracting informative qualitative signals from financial annual reports by training transformer-based classifiers on five related tasks (Relevance, Financial sentiment, Objectivity, Forward-looking, ESG) using multiple multi-task learning (MTL) strategies. The authors compare joint, sequential, weighted, ExGF-MTL, and TARS architectures, finding ExGF-MTL—the approach that explicitly feeds auxiliary-task predictions as features for the target task—provides the strongest gains, particularly for ESG detection. They validate the approach on a FTSE350 corpus with 2,651 annotated sentences and demonstrate that textual features derived from the ESG-focused MT models correlate with Reuters ESG scores, offering a pathway to link narrative style with quantitative ESG indicators. The work highlights the importance of task selection in MT settings, shows that joint training generally outperforms sequential approaches, and suggests that their feature-extraction pipeline can be extended to causal analysis and broader ESG-finance research.

Abstract

For assessing various performance indicators of companies, the focus is shifting from strictly financial (quantitative) publicly disclosed information to qualitative (textual) information. This textual data can provide valuable weak signals, for example through stylistic features, which can complement the quantitative data on financial performance or on Environmental, Social and Governance (ESG) criteria. In this work, we use various multi-task learning methods for financial text classification with the focus on financial sentiment, objectivity, forward-looking sentence prediction and ESG-content detection. We propose different methods to combine the information extracted from training jointly on different tasks; our best-performing method highlights the positive effect of explicitly adding auxiliary task predictions as features for the final target task during the multi-task training. Next, we use these classifiers to extract textual features from annual reports of FTSE350 companies and investigate the link between ESG quantitative scores and these features.