About Our Project

Welcome to our project on analyzing the sentiment of parliamentary speeches related to (non-)renewable energy sources. Our goal is to understand the role parliamentarians play in influencing their countries' transition to renewable energy by examining the sentiment expressed in their speeches over time. This research is crucial for understanding how political discourse shapes public policy and progress towards global climate goals.

Research Question & Motivation

Our research is driven by the urgent need to keep global warming well below 2 degrees Celsius. Achieving this requires a global shift from non-renewable to low-carbon, renewable energy sources. Policy support is crucial for accelerating this transition. However, in many countries, the level of support for energy policies varies significantly among parliamentarians. Some champion renewable energy projects, while others oppose them for various reasons. By analyzing the sentiment in parliamentarians' speeches related to different energy sources, we aim to better understand their influence on energy transition policies.

Our research question is: How does the sentiment in politicians’ speeches related to different (non-)renewable energy sources change over time? To answer this, we will analyze a dataset of parliamentary speeches from the UK 2009 and 2019. This analysis can then easily be extended to other countries, types of energy productions and years.

Dataset/Corpus

For our analysis, we utilize the ParlSpeech V2 database, which contains full-text data of 6.3 million parliamentary speeches from the key legislative chambers of Germany, Austria, Czech Republic, Netherlands, Spain, UK, Sweden, and Denmark. This publicly available dataset includes metadata such as date, speech number, speaker's full name, party affiliation, and the speech text. The UK House of Commons is particularly well-represented, while the Czech Poslanecká sněmovna Parlamentu has the least representation. The data is already organized and labeled in CSV format, ready for analysis.

Methods

Given the multilingual nature of the dataset, we will initially focus on English language texts from the UK. The text data will undergo preprocessing, including tokenization, standardization (lowercasing, removing stop-words and punctuation), and preparation into sentence-level documents. This cleaned corpus will be used to train a topic modeling algorithm.

We will employ a semi-supervised topic modeling approach, such as "guided LDA" from Gensim, using relevant seed-words to classify documents based on the distinct energy sources discussed. Additionally, we will use an open-source BERT-based model finetuned for sentiment analysis to assign a positive or negative sentiment to each speech.

Dicussion of Outcomes and Limitations

When looking at the data plotted as we have done, we can see some characteristics which stand out:

• most energy modes in dicussion appear to follow very similar trends with correlated spikes and troughs
• most values are quite low indicating not very high levels of "positiveness" for any given energy-mode

In the following, we suggest an explanation for each. Firstly, the similar trends may result from the fact that there are some years in which no parliamentary speeches from our dataset discussed energy-related matters which regular-expression based filtration could detect. In other years where matters of energy may have been a more popular topic of discussion, it is likely that the attention was received by nearly all the various modes.

Secondly, looking at the low values of positiveness, we suggest that this is a result of an unpolished approach to measure "how much a given speech talks about a particular energy-mode" involving the TF-IDF score of each mode of energy as a scaling factor. This score on its own indicates to us how important that given word (e.g., "coal", "solar", etc.) is to the documents in the given cluster, and these values are often small, in the magnitude of 10e-2. Therefore, likely because of this simplistic attempt at using this score as a proxy, we have ended up with values which are usually on the lower end. However without manual verification we cannot rule out the fact that there is a tendency in the UK parliament to generally speak about energy with a slightly negative sentiment on average because these topics may only come up when in times of crisis or when they are concerning another controversion socio-political discussion (such as climate change or oil deals with other countries, etc.).

Despite the limitations, we believe the model still captures some effect of energy economics on politicians and vice versa. For example, the UK's oil production rose substantially in 2018, with the UK Oil and Gas Authority promoting investments in new technologies and practices to extend the lifespan of existing fields and enhance production efficiency. This is reflected in our results in the peak of positive sentiment towards oil in the UK House of Commons speeches in 2018.

Dashboard and Future Work

For our first prototype and as a proof of concept, we conducted our sentiment analysis for the UK House of Commons, as it provides the highest number of speeches (around 2 million) over the longest time period (2009-2019). In the future, our methology can be extended to cover the other parliaments in the ParlSpeech v2 and potentially also the ParlSpeech v3 dataset.

References

Our work builds on established research in the field. Key references include:

• Dehler-Holland et al. (2022), who analyzed the legitimacy of wind power in German media using sentiment analysis and topic modeling.
• Rudowsky et al. (2013), who applied supervised machine learning to analyze the sentiment of parliamentary speeches in Austria.
• Abercrombie and Batista-Navarro (2020), who utilized BERT-based models for sentiment analysis on UK parliamentary speeches.

By leveraging these methodologies and the extensive ParlSpeech V2 database, our project aims to provide a comprehensive analysis of parliamentary sentiment towards energy policy, offering valuable insights into the political drivers of the global energy transition.