Papers by Minh Duc Bui
Multi3Hate: Multimodal, Multilingual, and Multicultural Hate Speech Detection with Vision–Language Models (2025.naacl-long)
Copied to clipboard
| Challenge: | a new study shows that cultural background significantly affects multimodal hate speech moderation models . a limited dataset excludes multi-modal forms of hate and excludes non-English-speaking cultures . the lowest pairwise label agreement between the USA and India is due to cultural factors . |
| Approach: | They use a multimodal and multilingual parallel hate speech dataset to examine cultural differences . they find that cultural background significantly affects multimodal hate speech annotation . |
| Outcome: | The proposed dataset shows that cultural background significantly affects multimodal hate speech annotation. |
Large Language Models Discriminate Against Speakers of German Dialects (2025.emnlp-main)
Copied to clipboard
| Challenge: | In Germany, more than 40% of the population speaks a regional dialect . however, dialect speakers face negative societal stereotypes . |
| Approach: | They construct a corpus that pairs sentences from seven regional German dialects with their standard German counterparts to assess their dialect usage bias. |
| Outcome: | The proposed model reproduces dialect usage bias in association task and decision task. |
Mind the Gap: A Closer Look at Tokenization for Multiple-Choice Question Answering with LLMs (2025.emnlp-main)
Copied to clipboard
| Challenge: | Recent studies have highlighted the significant performance variation that can arise from minor changes in prompt design. |
| Approach: | They propose to tokenize the space following the colon to facilitate automated answer extraction via next-token probabilities. |
| Outcome: | The proposed tokenization improves model calibration and improves confidence estimates. |
From If-Statements to ML Pipelines: Revisiting Bias in Code-Generation (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing methods to evaluate code generation bias focus on overt discrimination through simple conditional statements. |
| Approach: | They examine ML pipelines that exhibit substantially greater bias than simple conditionals . they challenge simple conditional statements as valid proxies for bias evaluation . |
| Outcome: | The proposed model underestimates real-world bias in generating machine learning pipelines . the model maintains equal performance on simple conditionals and ML pipelines, the study shows . |
On Generalization across Measurement Systems: LLMs Entail More Test-Time Compute for Underrepresented Cultures (2025.acl-long)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) should be able to provide accurate information irrespective of the measurement system at hand . |
| Approach: | They use newly compiled datasets to test if this is true for seven open-source LLMs. |
| Outcome: | The proposed model can provide accurate information regardless of the measurement system at hand. |