Predicting Roll-Call Voting from Parliamentary Speeches: Latent Topic Modelling and Network Analysis

The paper explores the possibilities of using parliamentary debates enriched with metadata for predicting individual MPs' positions in legislative voting. The ideological positions of MPs were estimated in two ways. The primary group of research methods consisted of generative topic-modelling methods, which infer latent topics (which the parliamentary speeches are considered to be mixtures of) based on word co-occurrence in utterances. We used two extensions of Latent Dirichlet Allocation (LDA) method: 1) the Structural Topic Model (STM; Roberts et al. 2016) which allows for incorporation of metadata into the topic model, assigning to each document its own prior topic distribution, basing on information provided by the metadata; and 2) Dynamic topic model (Blei, Laffery 2006) which allows for taking into account changes that occur in the semantic composition of the topic over time. As the complementary method, the network analysis method was used to calculate the frequency of specific MPs give speeches next or close in time to each other (temporal affinity). Both methods allow for the mapping of proximity of MPs on latent space (ideology) and for centrality detection (finding the most influential MPs whose behavior influence other MPs to speak in a similar way and time). For roll call analysis, the Poole’s (2000) Optimal Classification method was used to estimate positions of MPs. This methods allows for both perfect spatial voting and differences in party discipline, which is of great importance in Central and East European Parliaments. The study was conducted using data from the Lithuanian parliament (Kapočiūtė-Dzikienė, Šarkutė, Utka 2017) and the lower chamber of the Polish Parliament (Kwiatkowska 2019, forthcoming) collected within the cross-referencing database containing parliamentary speeches and roll call votes with meta data. The results prove that the MPs’ positions on latent topics are very good predictors of legislative voting and can be used to estimate the ideological positions of MPs. Statistically significant latent dimensions of debates indicate key topics of parliamentary competition and determine the positions of MPs in legislative voting. The analysis is further improved by including information from the network analysis, providing clusters emerging from the MPs’ speech temporal affinity.