RDP 2025-06: An AI-powered Tool for Central Bank Business Liaisons: Quantitative Indicators and On-demand Insights from Firms Appendix B: Spot-checking the Precision of the LM-based Topic Classifications

In this appendix, we perform a back-of-the-envelope assessment of the accuracy of all 14 topics classified by the transformer-based LM. To do this, we randomly sample 10 paragraphs from each topic, each with an LM-based topic probability of greater than 90 per cent, for a total of 140 paragraphs. We then assess the accuracy of these LM-based classifications allowing us to do a spot check of the LM's precision for each topic. The results are presented in Table B1 and show that topics that are more broadly defined (e.g. ‘costs’) or open to interpretation (e.g. ‘financing conditions’) – and so requiring more context and nuance to properly classify – are less precisely identified by the model.

Table B1: Spot Checking the Precision of the LM-based Topic Classifications
10 samples per topic
  Precision: TP / TP + FP (%)
Financing conditions 30
Costs 50
Non-labour costs 50
Prices 60
Employment 70
Margins 80
Investment or capex 80
Labour costs 90
Demand 90
Supply chains 90
Wages 100
Sales 100
Property or housing 100
Climate change 100
Overall 78

Sources: Authors' calculations; RBA.