Linking Economic Ideas and Narratives between Corpora
PS8-2
Presented by: Musashi Harukawa, Lucy Barnes
Recent works in political economy explore the importance of descriptive (Barnes forthcoming) and normative (Cavaillé 2021) ideas about the economy held by the public and elites (Killick 2021) for policy making and evaluation. Contrary to theoretical expectations, these mental models of the economy vary considerably (Barnes et al forthcoming) and can be explored systematically.
Using advances from computational linguistics in argument mining and text summarisation, we propose an automated framework for identifying and linking ideas about the economy expressed by politicians, the media and the public in multi-lingual and multi-media corpora consisting of elite interviews, surveys and newspapers from five countries.
To identify and structure ideas of the economy in text corpora, we adapt the parse-and-reduce approach of Ash et al (2021). Subject-verb-object (SVO) triples are extracted from the text using dependency parsers, then the space of SVO tokens is reduced using labelled WordNet associations and unsupervised word embedding clustering. Finally, we use semi-supervised label propagation and topic modelling methods to classify and cluster triples.
The main challenge results from our expectation that we do not expect to find identical clusters of ideas between politicians, the public and the media. Our key methodological contribution is a method for identifying and linking related ideas and narratives in disparate media and corpora. This is possible with descriptively intuitive and rich SVO triples, which allow for qualitative discursive comparison of the idea clusters. This tool has wider application for the empirical measurement of ideational flows.
Using advances from computational linguistics in argument mining and text summarisation, we propose an automated framework for identifying and linking ideas about the economy expressed by politicians, the media and the public in multi-lingual and multi-media corpora consisting of elite interviews, surveys and newspapers from five countries.
To identify and structure ideas of the economy in text corpora, we adapt the parse-and-reduce approach of Ash et al (2021). Subject-verb-object (SVO) triples are extracted from the text using dependency parsers, then the space of SVO tokens is reduced using labelled WordNet associations and unsupervised word embedding clustering. Finally, we use semi-supervised label propagation and topic modelling methods to classify and cluster triples.
The main challenge results from our expectation that we do not expect to find identical clusters of ideas between politicians, the public and the media. Our key methodological contribution is a method for identifying and linking related ideas and narratives in disparate media and corpora. This is possible with descriptively intuitive and rich SVO triples, which allow for qualitative discursive comparison of the idea clusters. This tool has wider application for the empirical measurement of ideational flows.