Papers by Zexue He
Detect and Perturb: Neutral Rewriting of Biased and Sensitive Text via Gradient-based Decoding (2021.findings-emnlp)
Copied to clipboard
| Challenge: | Written language carries explicit and implicit biases that can distract from meaningful signals; at worst they can lead to unfair outcomes. |
| Approach: | They propose a gradient-based rewriting framework that detects and perturbs sensitive components and regenerates fluent alternatives that are neutral in the sensitive attribute while maintaining the semantics of other attributes. |
| Outcome: | The proposed framework regenerates fluent alternatives that are neutral in the sensitive attribute while maintaining the semantics of other attributes. |
Synthetic Pre-Training Tasks for Neural Machine Translation (2023.findings-acl)
Copied to clipboard
| Challenge: | toxicity and bias can be addressed by pre-training with synthetic resources . BLEU scores are used to compare methods with real-world data . |
| Approach: | They propose several ways to generate obfuscated data from large parallel corpus and concatenating phrase pairs from small word-aligned corpus with synthetic parallel data without real human language corpora. |
| Outcome: | The proposed methods can be used to generate obfuscated data or synthetic parallel data without real human language corpora even with high levels of oblication. |
Controlling Bias Exposure for Fair Interpretable Predictions (2022.findings-emnlp)
Copied to clipboard
| Challenge: | Existing approaches to reduce bias in NLP tasks focus on protecting or isolating information related to a sensitive attribute, but they lack control over how much bias is required to be removed. |
| Approach: | They propose a favorable debiasing method that uses sensitive information ‘fairly’, rather than blindly eliminating it. |
| Outcome: | The proposed method achieves a trade-off between debiasing and task performance along with producing debiased rationales as evidence. |
MedEval: A Multi-Level, Multi-Task, and Multi-Domain Medical Benchmark for Language Model Evaluation (2023.emnlp-main)
Copied to clipboard
| Challenge: | Existing medical datasets require high quality domain-specific datasets. |
| Approach: | They propose a multi-level, multi-task, and multi-domain medical benchmark to facilitate the development of language models for healthcare. |
| Outcome: | The proposed model provides granular potential usage and supports a wide range of tasks. |
Cognitive Bias in Decision-Making with LLMs (2024.findings-emnlp)
Copied to clipboard
| Challenge: | Large language models inherit societal biases against protected groups and can be subject to functionally resembling cognitive bias. |
| Approach: | They propose a framework to uncover, evaluate, and mitigate cognitive bias in large language models by using a dataset containing 13,465 prompts to evaluate LLM decisions on different cognitive biases. |
| Outcome: | The proposed framework uncovers, evaluates, and mitigates cognitive bias in large language models, particularly in high-stakes decision-making tasks. |
Weakly Supervised Contrastive Learning for Chest X-Ray Report Generation (2021.findings-emnlp)
Copied to clipboard
| Challenge: | Radiology report generation aims at generating descriptive text from radiology images automatically. |
| Approach: | They propose a weakly supervised contrastive loss method that generates descriptive text from radiology images automatically. |
| Outcome: | The proposed method outperforms previous work on correctness and text generation metrics for two public benchmarks. |
InterFair: Debiasing with Natural Language Feedback for Fair Interpretable Predictions (2023.emnlp-main)
Copied to clipboard
| Challenge: | Debiasing methods in NLP models focus on isolating information related to a sensitive attribute (e.g., gender or race) but instead argue that a favorable debiaser should use sensitive information ‘fairly,’ with explanations, rather than blindly eliminating it. |
| Approach: | They propose that a favorable debiasing method should use sensitive information ‘fairly,’ with explanations, rather than blindly eliminating it. |
| Outcome: | The proposed approach reduces bias in explanations while maintaining the same prediction accuracy. |
Targeted Data Generation: Finding and Fixing Model Weaknesses (2023.acl-long)
Copied to clipboard
| Challenge: | Existing models fail systematically on specific subgroups of data, resulting in unfair outcomes and eroding user trust. |
| Approach: | They propose a framework that automatically identifies challenging subgroups and generates new data for those subgroup using large language models with a human in the loop. |
| Outcome: | The proposed framework improves accuracy on challenging subgroups while improving overall test accuracy. |
WildFeedback: Aligning LLMs With In-situ User Interactions And Feedback (2026.acl-long)
Copied to clipboard
Taiwei Shi, Zhuoer Wang, Longqi Yang, Ying-Chun Lin, Zexue He, Mengting Wan, Pei Zhou, Sujay Kumar Jauhar, Sihao Chen, Shan Xia, Hongfei Zhang, Jieyu Zhao, Xiaofeng Xu, Xia Song, Jennifer Neville
| Challenge: | Traditional alignment methods rely on human annotations and are subjective and misalignment with real-world user preferences. |
| Approach: | They propose a framework that leverages in-situ user feedback during conversations with LLMs to create preference datasets automatically. |
| Outcome: | The proposed framework identifies and classifies user feedback to LLM responses between conversation turns and creates examples of preferred and dispreferred responses according to user preferences. |
Leveraging Gloss Knowledge in Neural Word Sense Disambiguation by Hierarchical Co-Attention (D18-1)
Copied to clipboard
| Challenge: | Existing models for Word Sense Disambiguation use labeled data, but lack gloss knowledge. |
| Approach: | They propose a co-attention mechanism to generate co-dependent representations for context and gloss . they propose to incorporate gloss knowledge into neural networks for Word Sense Disambiguation . |
| Outcome: | The proposed model achieves state-of-the-art results on standard English all-words WSD datasets. |