Detecting context-based in-claim numerals in Financial Earnings Conference Calls

Abstract

Most investors tend to make decisions after analysing financial documents of organizations available online. These documents include financial reports, conversations, brochures, etc. While reading these documents investors need to ensure that they rely only on facts and do not get swayed away by claims which representatives of organizations make. Thus, it is essential to have an automated system for detecting whether numerals present in financial texts are in-claim. In this paper, we discuss a system for evaluating whether numerals present in financial texts are in-claim or out-of-claim. It is trained on the English version of the FinNum-3 corpus using two variants of the FinBERT model and a BERT model augmented with handcrafted features. Our best model, an ensemble of these 3 models, produces a Macro-F1 score of 0.8671 on the validation set and outperforms the existing baselines.