Analyzing political speeches presents a challenge: topic models identify themes but labeling them requires human effort. This study introduces transfer topic labeling, which uses domain-specific dictionaries to automate the process.
We tested this method on all UK House of Commons debates from 1935-2014—using CAP coding instructions—to show its potential for political science research. Our evaluation against expert coding revealed promising results:
• Method: Transfer topic modeling using domain-specific codebooks as a knowledge base.
• Performance: Simple to implement and compared well with human experts.
• Advantage: Outperformed neural network models in most cases.
This approach offers an accessible, replicable solution for automatically labeling themes.
| Transfer Learning for Topic Labeling: Analysis of the UK House of Commons Speeches 1935-2014 was authored by Alexander Herzog, Peter John, Slava Mikhaylov and Hannah Bechara. It was published by Sage in R&P in 2021. |
