A Greek Parliament Proceedings Dataset for Computational Linguistics and Political Analysis

Konstantina Dritsa, Aikaterini Thoma, Ioannis Pavlopoulos, Panos LouridasNeurIPS 2022

While large-scale diachronic (across-time) datasets of political discourse exist for "high-resource" languages like English or German, they remain rare for "resource-lean" languages like Greek. This data gap limits the ability of computational linguists and political scientists to study how language evolves within specific cultural and regional contexts.

Furthermore, detecting "semantic shifts"—how the meaning or usage of a word changes over time—is notoriously difficult. Different algorithms often produce inconsistent results even when run on the same data, a problem known as algorithmic instability. To tackle both data scarcity and methodological challenges, researchers have released a massive, 31-year dataset of Greek parliamentary proceedings (1989–2020). By benchmarking semantic shift algorithms on this data, the authors demonstrate how political language serves as a "living record" of historical events, such as the Greek economic crisis and shifting party coalitions.

The Greek Parliament Proceedings Dataset

The primary contribution of this work is a comprehensive, research-ready dataset transformed from raw parliamentary records.

  • Scale: 1,280,918 speeches from 5,355 sitting record files (2.12 GB).
  • Timeframe: July 1989 to July 2020.
  • Metadata: Includes speaker name, gender, political party, role (e.g., minister), and the specific parliamentary session/period.

To build this, the authors crawled the Greek Parliament website, converting over 5,000 PDFs, Word documents, and HTML files into raw text. They then used the Jaro-Winkler string similarity metric and auxiliary datasets of Greek names to resolve entities, accurately linking speeches to specific politicians despite nicknames, misspellings, or grammatical variations.

Methodological Insights: Tracking Semantic Shifts

The authors compared four state-of-the-art approaches to benchmark semantic shifts: Orthogonal Procrustes, Compass, Nearest Neighbors (NN), and Second-Order Similarity. They used the intersection@k metric to measure how much the results changed across multiple runs with different random seeds.

Stability of semantic shift detection approaches across different values of k.

The "Compass" Sweet Spot: The NN method showed the highest stability across all values of k. However, its results were often less "explainable" to human evaluators. The authors found that the Compass variation (which removes very frequent and very rare words) offered the best balance, performing nearly as well as NN on stability while providing much more meaningful linguistic insights.

Additionally, the study noted that "transitional" parliamentary periods (those lasting only a few months) have very low vocabulary overlap with other periods. If not handled carefully, this lack of overlap can lead to "false positive" semantic shifts.

Political and Social Insights

Beyond methodological benchmarks, the dataset provides a fascinating window into recent Greek history:

  • Language as a Crisis Mirror: The dataset vividly captures the 2008 economic crisis. For example, the word "haircut" (kourema) shifted from its literal meaning to referring to debt restructuring. Similarly, "brain" became inextricably linked to "brain drain" (fygí egkefálon) as skilled workers left the country. The word "refugee" (prosfygas) also shifted significantly around 2015, moving from a context of "cheap labor" to one reflecting the massive influx of asylum seekers.
  • Fluid Political Identities: By treating political party names as tokens, the authors showed how the "meaning" of a party shifts based on its alliances. For instance, the center-right New Democracy (ND) party's word embeddings moved closer to its rival PASOK during the years they shared a coalition government.
  • Gender Disparity in Discourse: The data reveals that while female representation is increasing, it remains low.
Percentage of female members and speakers across parliamentary periods by political party.

Interestingly, left-wing parties (such as SYN, KKE, and SYRIZA) generally show higher female participation and speech activity than right-wing parties, though no party has yet reached 50% parity.

Limitations and Future Work

The authors acknowledge that parliamentary language is highly formal and specific; it does not necessarily represent the everyday speech of the general Greek population.

Looking ahead, the researchers propose several exciting future directions. These include integrating the dataset with social media (like tweets from MPs) to create a more holistic model of political behavior, studying how new members of parliament assimilate their speech patterns to match their party leaders, and using the rich metadata to improve models that predict how a politician might vote based on their past rhetoric.

Summary updated

BibTeX
@inproceedings{dritsa2022greek,
  author = {Konstantina Dritsa and Aikaterini Thoma and Ioannis Pavlopoulos and Panos Louridas},
  title = {{A Greek Parliament Proceedings Dataset for Computational Linguistics and Political Analysis}},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {2022},
}
Figure