Political Stance in Danish

Rasmus Lehmann, Leon DerczynskiNODALIDA 2019 · OpenReview

Abstract

The task of stance detection consists of classifying the opinion within a text towards some target. This paper seeks to generate a dataset of quotes from Danish politicians, label this dataset to allow the task of stance detection to be performed, and present annotation guidelines to allow further expansion of the generated dataset. Furthermore, three models based on an LSTM architecture are designed, implemented and optimized to perform the task of stance detection for the generated dataset. Experiments are performed using conditionality and bi-directionality for these models, and using either singular word embeddings or averaged word embeddings for an entire quote, to determine the optimal model design. The simplest model design, applying neither conditionality or bi-directionality, and averaged word embeddings across quotes, yields the strongest results. Furthermore, it was found that inclusion of the quotes politician, and the party affiliation of the quoted politician, greatly improved performance of the strongest model.

Introduction

In the modern digital landscape, the volume of political discourse makes manual analysis increasingly difficult. Stance detection—the automated classification of whether a text is for, against, or neutral toward a specific target—provides a quantitative lens for understanding political leanings. While well-established for English, Danish remains a lower-resource language in NLP. This research addresses the challenge of building robust stance detection models for Danish politics, specifically focusing on the nuanced and often implicit nature of political rhetoric.

Methodology

The study compared two primary deep learning architectures to determine the most effective approach for low-resource stance detection:

  1. Conditional LSTM: A sequential model where the initial hidden state is conditioned on a vector representing the speaker's identity and party.
  2. Multi-Layered Perceptron (MLP): A feed-forward network that processes the average of word embeddings for a given quote, concatenated with speaker metadata.

Both models utilized 300-dimensional fastText embeddings. A critical component of the model design was the inclusion of contextual metadata—one-hot encodings for the politician and their party affiliation—to provide the model with "priors" on the speaker's likely stance.

The MLP architecture used in the study, highlighting the integration of word embeddings with speaker and party metadata.

Surprisingly, the simpler MLP architecture outperformed the more complex LSTM. In scenarios with limited data, complex models like LSTMs often struggle with overfitting or failing to capture long-range dependencies, whereas simple averaging of embeddings proves more robust.

Dataset and Annotation

A novel dataset was curated from the Danish news agency Ritzau, comprising 898 quotes from 63 politicians across 9 parliamentary parties. The dataset focuses on immigration policy, categorized into national and centralized (EU-level) issues.

Stance distribution across the Danish political spectrum, showing a significant skew toward 'For' stances, particularly among right-wing parties.

The dataset exhibits a notable skew: 57.2% of quotes are labeled as for tighter immigration restrictions. This reflects the source material's tendency to quote right-wing parties more frequently on this specific topic.

Results and Key Insights

The MLP model achieved the highest performance with an $F1_{macro}$ of 0.575 and an $F1_{micro}$ of 0.717. Ablation studies revealed that the inclusion of metadata was the single most important factor for success. Without speaker and party information, the MLP's $F1_{macro}$ dropped to 0.138.

This finding underscores that in political discourse, "context is king." Stance is often expressed through proxy language—such as referring to immigrants as "resources" versus "nuisances"—rather than explicit declarations. Knowing the speaker's identity allows the model to correctly interpret these linguistic nuances.

Conclusion and Future Work

While the simple MLP provided a strong baseline, its performance is constrained by the dataset size (898 quotes). Future research should focus on expanding the dataset to enable the use of more sophisticated architectures like Transformers. Additionally, addressing the label imbalance and exploring cross-target stance detection (e.g., training on immigration and testing on climate policy) remain important avenues for improving the generalizability of Danish political stance detection.

Summary updated

Abstract

The task of stance detection consists of classifying the opinion within a text towards some target. This paper seeks to generate a dataset of quotes from Danish politicians, label this dataset to allow the task of stance detection to be performed, and present annotation guidelines to allow further expansion of the generated dataset. Furthermore, three models based on an LSTM architecture are designed, implemented and optimized to perform the task of stance detection for the generated dataset. Experiments are performed using conditionality and bi-directionality for these models, and using either singular word embeddings or averaged word embeddings for an entire quote, to determine the optimal model design. The simplest model design, applying neither conditionality or bi-directionality, and averaged word embeddings across quotes, yields the strongest results. Furthermore, it was found that inclusion of the quotes politician, and the party affiliation of the quoted politician, greatly improved performance of the strongest model.
BibTeX
@inproceedings{lehmann2019political,
  author = {Rasmus Lehmann and Leon Derczynski},
  title = {{Political Stance in Danish}},
  booktitle = {NODALIDA},
  year = {2019},
  url = {https://openreview.net/forum?id=HqwP2rwSB6_},
}
Figure