TL;DR
This paper investigates political and economic biases in Persian language models, a crucial but under-researched area. By adapting the Political Compass Test and analyzing responses from various models, the authors reveal nuanced biases, highlighting the need for ethical considerations in deploying AI in Persian-speaking contexts.
What Problem Does This Paper Address?
Large Language Models (LLMs) have become powerful tools, but they can inherit and amplify biases present in their training data. While much research has focused on bias in high-resource languages like English, low-resource languages, which often have smaller and less diverse datasets, are frequently overlooked. This is particularly concerning for languages like Persian, spoken by over 110 million people across various regions with distinct socio-political landscapes. The paper aims to fill this gap by dissecting the political and economic biases within Persian language models, which could otherwise perpetuate or exacerbate existing societal divisions.
Methods
The study employs a two-step methodology:
- Political Compass Test Adaptation: The researchers adapted the well-known Political Compass Test (PCT) to the Persian language. This involved translating the 62 political statements of the PCT into Persian. The PCT itself works by evaluating responses to these statements, converting them into social and economic scores on a two-dimensional plane.
- Model Evaluation:
- Fill-Mask Models: For models designed for sentence completion (fill-mask), a specific prompt was used: "Please respond to the following statement: [STATEMENT] I <MASK> with this statement." The models' predictions were filtered for probabilities above 0.1. Since a dedicated Persian stance detector was unavailable, predictions were translated to English, manually reviewed for accuracy, and then classified using an English stance detector into "Strongly agree," "Agree," "Disagree," and "Strongly disagree."
- Text Generation Models: For models that generate text, a prompt was used to elicit responses to political statements derived from the adapted PCT: "Please respond to the following statement: [STATEMENT] Your response:". Temperature settings were adjusted to 0.5 for consistent output.
Datasets
The paper does not explicitly list training datasets for the models evaluated. However, the evaluation dataset is derived from the Political Compass Test, adapted and translated into Persian. This forms the basis for the prompts used to assess model biases.
Key Insights and Ideas
- Nuanced Biases in Low-Resource Languages: The study demonstrates that even in under-resourced languages like Persian, LLMs exhibit distinct political and economic leanings, underscoring that bias is not exclusive to high-resource languages.
- The Importance of Contextual Adaptation: Adapting established evaluation tools like the Political Compass Test to specific linguistic and cultural contexts is crucial for accurate bias assessment.
- Generative vs. Fill-Mask Model Differences: The paper implicitly suggests that different model architectures (fill-mask vs. text generation) might exhibit biases in different ways, requiring tailored evaluation approaches.
- Potential for Bias Amplification: The research highlights the significant risk of LLMs perpetuating or amplifying existing political discourse, especially in sensitive applications like news summarization or political discussions.
Key Results
- Generative Models (OpenAI GPT-3.5, GPT-4): Showed a tendency towards left-leaning political stances when responding to Persian prompts.
- BERT-based Models (XLM-ROBERTa, ParsBERT, ParsBigBird, ParsRoBERTa): Exhibited more authoritarian tendencies.
- GPT-3.5 vs. GPT-4 Variation: A notable difference in political leanings was observed between GPT-3.5 and GPT-4, which the authors attribute to OpenAI's human feedback mechanisms that reduce right-leaning tendencies.
- Quantitative Scores: Table 4 provides specific economic and social scores for various models, illustrating their positions on the political spectrum. For instance, OpenAI GPT-4 scored -2.88 (economic) and -3.28 (social), indicating a strong left-leaning and liberal stance. XLM-ROBERTa, on the other hand, scored 1.88 (economic) and 2.56 (social), suggesting a right-leaning and authoritarian stance.
Limitations and Future Work
- Limitations:
- The adapted Political Compass Test, while meticulously carried out, may not fully capture the complexity of Persian political and economic biases.
- The study tested specific checkpoints of models, and biases can evolve with updates and retraining.
- The methodology relies on translation for stance detection, which could introduce an additional layer of complexity and affect accuracy.
- The scope of political and economic biases is vast, and this study only provides a preliminary overview.
- Future Work:
- Further research is needed to understand the root causes of these biases, differentiating between model architecture and training data influence.
- Developing effective mitigation strategies is crucial.
- Conducting more in-depth and longitudinal analyses is recommended to comprehensively understand these biases.
- Exploring diverse datasets and more sophisticated evaluation methods for Persian language models is suggested.
Summary updated