Word Embeddings: Providing More Meaningful Representations of Legislators

• In Event: New Methods for Text-as-Data

Abstract

Research on legislative behavior has long been interested in accurately explaining drivers of legislative behavior. Why does a member of parliament (MP) choose to vote alongside their party, support a specific bill or argue in favor a specific policy? What are the interests and agenda that drive the MPs' actions? To quantitatively explain MP’s legislative choices, compare their interests and agendas, a quantitative and precise representation of MPs is needed. Often, these representations are described as positions in a latent political space as they allow for quantitative comparisons between MPs. So far, methods exist to use voting, cosponsorship or speech data to estimate these positions of MPs. However, they suffer from significant limitations. For instance, voting records are generally biased by party discipline. MPs’ personal relationships bias cosponsorship data. One of the data sources with the least limitations are text data, for instance from political speeches or proposed bills.

Traditional methods for estimating political positions from text rely on word counts. However, they do not account for word context and cannot be used in multilingual settings. To better process and understand texts, computer scientists have developed a technique called word embeddings, which represent a numerical representation of a word in a high-dimensional space. It captures the semantic and syntactic relationships between words in a continuous vector space, making it easier to process and relate to other words. How can word embeddings be used to elicit positions or representations of MPs from texts?

In this paper we show that word embeddings of legislative speeches and bills allow for more holistic representations of MPs. We show how word embeddings can be calculated for MPs and examine which socio-economic factors they capture in order to present a more holistic representation of MPs. Furthermore, we show that by training these embeddings on an ideology-related task (voting agreements), these embeddings can be used as a proxy for political positions. When comparing these political positions to value-based survey answers, we find that they far-outperform voting-based ideology scores.

In our study, we use Sentence BERT, a powerful model that uses tokenization, predefined BERT models, and a pooling layer to understand and represent the meaning of text. It outputs 512-dimensional vector representations of each sentence, allowing for more efficient and accurate representations of texts.

We focus on the Swiss parliament as it represents both a multi-party system (currently consisting of six parliamentary groups) as well as multilingual system (where four official languages are spoken on the parliamentary floor). We use data from four legislative periods (47-50th, 30.11.2003-01.12.2019). A total of 23'742 legislative bills were introduced and 74'471 speeches were held in parliament by MPs as individuals (i.e., not as speakers of committees)
By embedding the complete set of bills and speeches of each MP, we create a fixed-sized vectoral representation for each MP. We then proceed to ask two important questions:
First, what do these embeddings of MPs record? We show that these vectoral representations hold copious information. Not only do these embeddings hold information related to an MP’s parliamentary group affiliation and voting decisions. We demonstrate that the embeddings also hold information on other key characteristics of MPs, such as their gender, the region represented, their education level, and even their job category. This indicates that MP embeddings reflect holistic representations of MPs.
Second, we ask: are these embeddings useful for a specific task? Can they be sharpened to one specific aspect of an MP? To answer this, we chose to try and encode ideology of MPs more closely into these MP embeddings. To refine the embeddings to map ideology more closely, we train them with additional data. We show that by training our embeddings on voting data, the resulting representations of MPs cluster together by parliamentary group if we look at a two-dimensional space. Furthermore, our trained representations can explain ideology-related survey questions with great accuracy.

We show that legislative speech and bill text embeddings make for holistic representations of MPs. Further, we show that by training these embeddings on voting data to focus on ideology, our MP representations become interpretable and comparable.
Moving away from existing limited methods for eliciting representations of MP, embeddings open the door for new ways of encoding MP’s behavior, ideology and social standing. By refining them on a specific task, we can (i) project positions of MPs where we have speech data, but no survey data; and (ii) we can elicit positions of MPs not included in our training sample. Our approach allows us to create a more holistic representation of an MP, that goes far beyond an MPs' party affiliation or voting behavior.

Authors

• Sophia Schlosser

• Laurence Brandenberger