Papers by Khalil Mrini
Rethinking Self-Attention: Towards Interpretability in Neural Parsing (2020.findings-emnlp)
Copied to clipboard
| Challenge: | Recent work shows that attention mechanisms provide arguably explainable attention distributions that can help to interpret predictions. |
| Approach: | They propose a new self-attention layer where attention heads represent labels. |
| Outcome: | The proposed model obtains state-of-the-art results on the Penn Treebank and Chinese Treebank. |
Medical Question Understanding and Answering with Knowledge Grounding and Semantic Self-Supervision (2022.coling-1)
Copied to clipboard
Khalil Mrini, Harpreet Singh, Franck Dernoncourt, Seunghyun Yoon, Trung Bui, Walter W. Chang, Emilia Farcas, Ndapa Nakashole
| Challenge: | Current medical question answering systems have difficulty processing long, detailed and informally worded questions . a growing number of approaches attempt to enhance the processing of consumer health questions - or medical question understanding . |
| Approach: | They propose a medical question understanding and answering system with knowledge grounding and semantic self-supervision that matches a user question with a trusted medical knowledge base and retrieves a fixed number of relevant sentences from the corresponding answer document. |
| Outcome: | The proposed system retrieves more relevant answers while achieving 20 times faster. |
Recursive Tree-Structured Self-Attention for Answer Sentence Selection (2021.acl-long)
Copied to clipboard
| Challenge: | Recent top-performing models in Answer Sentence Selection use self-attention and transfer learning, but not syntactic structure. |
| Approach: | They propose a recursive, tree-structured self-attention model that can represent all levels of syntactic parse trees with only one additional layer. |
| Outcome: | The proposed model can represent all levels of syntactic parse trees with only one additional layer without transfer learning. |
Detection, Disambiguation, Re-ranking: Autoregressive Entity Linking as a Multi-Task Problem (2022.findings-acl)
Copied to clipboard
| Challenge: | Existing methods for entity linking do not use a knowledge base or candidate sets. |
| Approach: | They propose an autoregressive entity linking model that is trained with two auxiliary tasks and learns to re-rank generated samples at inference time. |
| Outcome: | The proposed model improves on two biomedical datasets and a news domain dataset without the use of a knowledge base or candidate sets. |
Sentence-level Privacy for Document Embeddings (2022.acl-long)
Copied to clipboard
| Challenge: | Existing studies on the definitions of good privacy for natural language use argue that different applications and models require different definitions. |
| Approach: | They propose a technique that uses a combination of statistics and language modeling to produce high (768) dimensional, general -SentDP document embeddings that guarantee a single sentence can be substituted with any other sentence. |
| Outcome: | The proposed method outperforms baseline methods with weaker guarantees like word-level Metric DP and outperformed baseline methods. |
A Gradually Soft Multi-Task and Data-Augmented Approach to Medical Question Understanding (2021.acl-long)
Copied to clipboard
Khalil Mrini, Franck Dernoncourt, Seunghyun Yoon, Trung Bui, Walter Chang, Emilia Farcas, Ndapa Nakashole
| Challenge: | Existing methods for medical question understanding often fail to provide high recall in answer retrieval. |
| Approach: | They propose a multi-task learning method with data augmentation for medical question understanding that uses just one dataset to optimize for both tasks. |
| Outcome: | The proposed method outperforms existing MTL methods across 4 datasets of medical question pairs in ROUGE scores, RQE accuracy and human evaluation. |
Agent-Testing Agent: A Meta-Agent for Automated Testing and Evaluation of Conversational AI Agents (2026.eacl-long)
Copied to clipboard
| Challenge: | Robust, developer-friendly evaluation remains a bottleneck. |
| Approach: | They propose a meta-agent that combines static code analysis, developer interrogation, literature mining, and persona-driven adversarial test generation whose difficulty adapts via judge feedback. |
| Outcome: | The agent-testing agent (ATA) surfaces more diverse and severe failures than expert annotators while matching severity, and finishes in 20β30 minutes versus ten-annotator rounds that took days. |