Financial Text Categorisation with FinBERT on Key Audit Matters

Accurate extraction and categorisation of large amount of textual data is pivotal for efficient decision-making in this rapidly changing domain of finance. In this technical paper, the combination of natural language processing (NLP) and machine learning techniques was introduced. This novel approach was specifically developed to extract unstructured financial text data from the annual reports of public listed companies in Bursa Malaysia and categorise relevant topics of Key Audit Matters (KAM). In particular, areas of interest were extracted in the financial text format, and the integration of the state-of-the-art model enabled the categorisation of the extracted content into relevant topics. Optical character recognition (OCR) tool was first utilised for text pre-processing, which involved converting PDF files to a structured data format. For feature extraction, bounding box coordinates were identified and then fed into the areas of interest prediction model. Following that, t-Distributed Stochastic Neighbour Embedding (t-SNE) was performed to establish and optimise KAM categories. With the accuracy of 94% for KAM topic categorisation, the FinBERT pre-trained model demonstrated better performance than that of MiniLM.