💻 computer science

Investigating Bias in Bulgarian in the Context of Large Language Models

This paper presents a manually annotated Bulgarian dataset to evaluate bias detection, revealing high agreement among human annotators but significant divergence between humans and large language models, while demonstrating that Bulgarian-language prompting improves model alignment with human judgments.

Original authors: Ivelina Stoyanova

Published 2026-08-10

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Ivelina Stoyanova

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are teaching a super-smart robot how to understand human stories. You give it millions of books, articles, and websites to read, hoping it learns not just grammar, but also how to be fair, kind, and neutral. This is the world of Large Language Models (LLMs), the AI brains behind chatbots and search engines. But here's the catch: robots learn by mimicking what they see. If the books they read contain old-fashioned stereotypes or unfair opinions about certain groups of people, the robot might accidentally learn those biases too. It's like a student who only reads history books written by one side of a war; they might think that side is always right. Scientists are worried because if these robots start making decisions about jobs, loans, or news, they could accidentally treat people unfairly. The big question is: Can we teach these robots to spot unfairness in text, and do they see it the same way humans do?

This paper dives into that question, but with a twist: it looks at the Bulgarian language. Most AI research focuses on English, which is like studying only one flavor of ice cream and assuming you understand the whole freezer. The researchers wanted to see if the same rules apply to a language with fewer digital resources and a very different culture. They built a special "test kit" using 3,177 sentences from Bulgarian Wikipedia. Two human experts read these sentences and graded them on a scale of 0 to 5 to see how biased they were, looking for unfairness related to gender, religion, race, appearance, and disability. Then, they asked three different AI models (Gemma 4, BgGPT-Gemma-3, and Qwen 3) to grade the exact same sentences.

The results were a bit of a shock, like asking a human and a robot to judge a magic trick and getting completely different answers. The two humans agreed with each other almost all the time (only 4.9% of the time they disagreed on whether a sentence was biased). However, when the AI models tried to do the same job, they were much less reliable. The robots disagreed with the humans 24% to 34% of the time. Even worse, when the humans said a sentence was biased, the robots often didn't see it, and when the robots did flag something as biased, the humans often thought it was fine.

The paper suggests that the robots aren't just "wrong"; they are playing by a different rulebook. The humans were very strict about how bad the bias was (giving high scores for strong bias), but the robots were much more cautious, giving very low scores even when humans saw strong unfairness. It's as if the humans are shouting, "That's terrible!" while the robots are whispering, "Maybe a little rude?"

There was one interesting discovery, though. One of the AI models, BgGPT-Gemma-3, was built specifically for Bulgaria. When the researchers asked it questions in English, it did a poor job. But when they switched the instructions to Bulgarian, the robot suddenly started seeing bias more like a human did. It's as if speaking the robot's native language woke up a different part of its brain, helping it understand the cultural context better.

The study concludes that we can't just trust robots to fix bias in low-resource languages like Bulgarian yet. They are currently too different from human judgment. The author suggests that instead of relying on a single AI to do the grading, we might need to use a team of different AIs and then have humans check their work. They also emphasize that we need to build better dictionaries and tools specifically for Bulgarian culture, because the robots are currently missing the subtle, cultural clues that humans pick up on instantly. Until then, the robots are still learning how to be truly fair.

Technical Summary: Investigating Bias in Bulgarian in the Context of Large Language Models

Problem Statement
Large Language Models (LLMs) frequently reflect and amplify social biases present in their training data. While significant research exists for high-resource languages like English, there is a critical gap in bias detection methodologies, datasets, and lexical resources for low-resource languages, specifically Bulgarian. The paper addresses the lack of a unified definition of bias applicable to both humans and LLMs, the scarcity of Bulgarian-specific bias lexicons, and the uncertainty regarding whether LLMs can reliably align with human judgments on bias perception in culturally specific contexts. The study poses three core questions: Can an operational definition of bias ensure consistency for both humans and LLMs? Do human annotators agree on bias perception? Do LLMs align with human judgments, and if not, how and why do they diverge?

Methodology
The study employs a multi-stage approach combining lexical resource construction, dataset creation, manual annotation, and LLM evaluation.

Lexical Resource Construction: A "Bulgarian Bias Dictionary" was manually compiled and expanded using the Bulgarian WordNet and SentiWordNet. It contains 3,786 entries categorized into five bias types (Gender, Religion, Race/Ethnicity, Physical Appearance, Disability) plus a general category. Entries include bias signal words and bias-marking words, annotated with attributes such as part of speech, bias polarity, register (e.g., derogatory, colloquial), and evaluative strength.
Dataset Creation: A dataset of 3,177 sentences was extracted from Bulgarian Wikipedia (articles and discussion pages) dated December 2025. Sentences were preprocessed (tokenized, lemmatized) and initially scored using a lexicon-based algorithm. This algorithm calculates a sentence-level score based on the evaluative strength, polarity, and proximity of bias signal words and bias-marking words within a 10-token window.
Manual Annotation: Two native Bulgarian speakers manually annotated the sentences for the presence and intensity of bias on an ordinal scale from 0 (no bias) to 5 (strong bias) for each of the five bias types and an "Other" category. Inter-Annotator Agreement (IAA) was measured using weighted Cohen's $\kappa$ , Krippendorff's $\alpha$ , and the Intraclass Correlation Coefficient (ICC).
LLM Evaluation: Three LLMs were evaluated against the human annotations:
- Gemma 4: A multilingual instruction-tuned model with Bulgarian in its pre-training corpus.
- BgGPT-Gemma-3: A Bulgarian-adapted model (12B parameters) based on Gemma 3, tested with prompts in both English and Bulgarian.
- Qwen 3: A multilingual model (8B parameters) with Slavic language coverage but no explicit Bulgarian fine-tuning.
  The models were tasked with detecting bias and assigning intensity scores. Agreement between humans and LLMs was evaluated using the same metrics as human-human agreement.

Key Contributions

Bias Annotated Dataset for Bulgarian: The creation of a manually annotated dataset of 3,177 sentences covering five specific bias types, serving as a benchmark for low-resource language bias detection.
Bulgarian Bias Dictionary: A structured lexical resource of 3,786 entries designed to identify bias signals and evaluative markers specific to Bulgarian.
Operational Definition of Bias: A working definition of sentence-level bias that accounts for systematic non-neutral attitudes, applicable to both human and machine evaluation.
Empirical Analysis of Human-LLM Alignment: A comprehensive evaluation of how well current LLMs align with human judgments in a low-resource, culturally specific setting, revealing significant divergences in both detection and intensity scoring.

Results

Human-Human Agreement: Annotators showed high agreement on the binary classification of bias (biased vs. non-biased), with a mismatch rate of only 4.9% and a weighted $\kappa$ of 0.863. However, agreement on intensity scoring (0–5 scale) was lower, particularly for non-zero cases, where Krippendorff's $\alpha$ averaged -0.160, indicating disagreement greater than chance on specific intensity scores.
Human-LLM Alignment: All LLMs showed substantial divergence from human judgments.
- Mismatch Rates: LLMs exhibited mismatch rates of 24% to 34% on the binary task, compared to 4.9% for humans.
- Agreement Metrics: Weighted $\kappa$ values between humans and LLMs ranged from 0.108 (Qwen 3) to 0.255 (Gemma 4), falling in the slight-to-fair range.
- Intensity Scoring: LLMs systematically assigned lower intensity scores than humans. Non-zero agreement metrics ( $\alpha_{nz}$ ) were uniformly negative across all models, indicating that when LLMs flagged bias, they often flagged different sentences or assigned different severity levels than humans.
Model-Specific Findings:
- Prompt Language: Using Bulgarian prompts for BgGPT-Gemma-3 improved alignment with human judgments (increasing biased overlap from 12.8% to 21.6% and reducing mismatch rates) compared to English prompts.
- Bias Type Variability: LLMs performed relatively better on "Appearance" and "Disability" categories (likely due to explicit lexical markers) but performed poorly on "Gender" and "Religion," which rely on subtle cultural and contextual cues.
- Model Clustering: Gemma-family models (Gemma 4 and BgGPT variants) clustered together, suggesting shared biases from their pre-training lineage, while Qwen 3 formed a distinct cluster with the lowest agreement.

Significance and Claims
The paper claims that current LLMs are unreliable as standalone substitutes for human annotators in bias detection, particularly for low-resource languages and culturally specific contexts. The findings underscore that:

Bias Perception is Culturally Situated: LLMs struggle with biases that require pragmatic and cultural competence (e.g., gender roles, religious practices), whereas they perform better on lexically explicit biases.
Prompt Language Matters: Target-language prompting significantly improves LLM alignment with human cultural expectations, suggesting that prompt language is a critical experimental variable, not just an implementation detail.
Need for Grounded Resources: The study highlights the necessity of culturally grounded lexical resources and annotation methodologies for low-resource languages, as transfer learning from multilingual models is insufficient for capturing specific bias patterns.
Methodological Caution: High inter-LLM agreement should not be interpreted as a proxy for annotation quality, as models may share systematic biases from their training data rather than converging on human social judgments.

The author concludes that while LLMs can partially replicate binary bias detection, they diverge significantly in intensity scoring and category sensitivity. Future work should focus on fine-tuning with target-language bias corpora, integrating cultural context into prompts, and expanding datasets to diverse domains to improve fairness and performance in low-resource settings.

Technical Summary: Investigating Bias in Bulgarian in the Context of Large Language Models

More like this