TopicShoal: Scaling Partisanship Using Semantic Search

TopicShoal: Scaling Partisanship Using Semantic Search Sami Diaf , Ulrich Fritsche Department of Socioeconomics Universit¨at Hamburg [sami.diaf,ulrich.fritsche]@uni-hamburg.de Abstract Document scaling techniques have been widely used in political science to infer parti- sanship measures and to rank documents on a scale of ideal points, based on bag-of-word ap- proaches. These approaches typically under- estimate the semantic and syntactic patterns contained in the corpus. Recent advances in natural language processing, particularly se- mantic search models, offer an improved topic coherence due to a semantic space of em- bedded words and documents, whose struc- ture is able to identify topics without setting their number as a hyperparameter. We pro- pose a scaling technique, namely TopicShoal, that extracts meaningful topic vectors using a semantic search technique (Top2Vec) and scales partisanship among speakers or par- ties using a Bayesian factor analysis on the document-topic distances, thereby enabling a semantic explanation of the ideal points’ varia- tions. This novelty, suited for both monolin- gual and multilingual corpora, addresses the bag-of-word constraint by capturing the nar- rative signals in the corpus and exploiting a coherent and independent topic vector struc- ture. Applied to a corpus of German party manifestos and Deutsche Bundesbank execu- tive board members’ speeches, TopicShoal suc- cessfully identifies discourse-level differences among parties and speakers via topic inten- sities, whose projection on the ideal points’ scale reveals common debated themes and other sideline interests that differentiate parties and speakers. 1 Introduction Text mining in political science comprises distinct families of methods usually applied to monolingual text data. Topic models define probabilistic models used to extract groups of words with a semantic meaning, referred to as topics based on a generative model of texts, while the document scaling family gathers probabilistic as well as non-probabilistic approaches used to infer a unidimensional scale as- sumed to be a proxy of ideal-points or (ideological) positions prevailing among speakers or parties. Non-probabilistic scaling techniques are based on pre-established wordlists from reference texts (Laver et al., 2003) whose availability outside the English language is limited, while probabilistic techniques are mostly based on the assumption of a Poisson distribution for word frequencies, as for Wordfish (Slapin and Proksch, 2008) which in- fers a unidimensional, normally distributed N(0, 1) scale for document positions, or the Poisson re- duced rank models which permit to endow a time- variability to the learned scale (Jentsch et al., 2020). Wordshoal (Lauderdale and Herzog, 2016) uses Wordfish estimates over distinct debates to aggre- gate the results at the level of speakers, where dif- ferences in document positions within debates ap- proximate the ideological stance between speak- ers. Such schemes have been used in political sci- ences to measure polarization of political parties in the United Kingdom (Goet, 2019), investigate left-right differences (D¨aubler and Benoit, 2021), in Germany for parties’ manifestos (Jentsch et al., 2021) or for economic institutions’ forecasting re- ports (Diaf et al., 2022) and were found to have some drawbacks in applications with small cor- pora or limited vocabulary (Hjorth et al., 2015) and to …

Content truncated to 500 words