Measuring Topics Using Cross-Domain Supervised Learning: Methods and Application to New Zealand Parliament
Moritz Osnabrügge, Elliott Ash, Massimo Morelli · Repository for Publications and Research Data (ETH Zurich) · 2020
Not cached yet. This record was fetched live for this request and is stored after the lookup. Reopening the page should serve the cached copy.
Repository for Publications and Research Data (ETH Zurich) · ETH Zurich · report · 2020
Abstract
This paper studies and assesses a novel method for assigning topics to political texts: cross-domain supervised learning.A machine learning algorithm is trained to classify topics in a labeled source corpus and then applied to extrapolate topics in an unlabeled target corpus.An advantage of the method is that, unlike standard (unsupervised) topic models, the set of assigned topics are interpretable and scientifically meaningful by construction.We demonstrate the method in the case of labeled party manifestos (source corpus) and unlabeled parliamentary speeches (target corpus).Besides the standard cross-validated within-domain error metrics, we further validate the crossdomain performance by labeling a subset of target corpus documents.We find that the classifier assigns topics accurately in the parliamentary speeches, although accuracy varies substantially by topic.To assess the construct validity, we analyze the impact on parliamentary speech topics of New Zealand's 1996 electoral reform, which replaced a first-past-the-post system with proportional representation.
Open the resolved locator · 10.3929/ethz-b-000414595
Metadata licence: OpenAlex metadata (CC0)