Scaling Arabic Medical Chatbots Using Synthetic Data: Enhancing Generative AI with Synthetic Patient Records
The development of medical chatbots in Arabic is significantly constrained by the scarcity of large-scale, high-quality annotated datasets. This study proposes a scalable synthetic data augmentation strategy to expand the training corpus to 100,000 records, enhancing the capabilities of large language models in Arabic healthcare applications.
Binary Quantization For LLMs Through Dynamic Grouping
This research proposes a novel optimization objective for binary quantization, enhancing blocked quantization by dynamically identifying optimal unstructured sub-matrices. Experimental results demonstrate significant reductions in model size without sacrificing performance, making it a viable solution for resource-constrained applications.
Are LLMs Better than Reported? Detecting Label Errors and Mitigating Their Effect on Model Performance
This study explores the efficacy of existing NLP benchmarks, highlighting the prevalence of label errors that distort the perceived performance of LLMs. By implementing LLMs to detect these errors, the research improves overall model accuracy and provides implications for data quality in NLP tasks.
FinMTEB: Finance Massive Text Embedding Benchmark
Introducing the Finance Massive Text Embedding Benchmark (FinMTEB), tailored for the financial domain with 64 datasets across various tasks. This benchmark aims to facilitate the assessment and advancement of embedding models specifically for financial applications, revealing critical insights regarding model performance.
SPECS: Specificity-Enhanced CLIP-Score for Long Image Caption Evaluation
SPECS is a reference-free metric that enhances the evaluation of long image captions by emphasizing specificity. It demonstrates efficiency comparable to LLM-based metrics and is practical for iterative evaluations, providing valuable insights for model development in image captioning tasks.
A Topic Modeling Analysis of Stigma Dimensions, Social, and Related Behavioral Circumstances in Clinical Notes Among Patients with HIV
This paper characterizes stigma dimensions affecting people living with HIV through NLP methods applied to clinical notes. The findings suggest actionable insights for improving HIV care outcomes by utilizing automated techniques to assess stigma in real-world documentation contexts.
Too Helpful, Too Harmless, Too Honest or Just Right?
This study proposes TrinityX, a modular alignment framework that utilizes a Mixture of Calibrated Experts to improve language model outputs in terms of Helpfulness, Harmlessness, and Honesty. The framework achieves significant performance improvements while reducing resource consumption.
Interdisciplinary Research in Conversation: A Case Study in Computational Morphology for Language Documentation
This position paper advocates for incorporating user-centered design in computational morphology research to bridge the gap between research and practical application in language documentation. The findings highlight necessary research directions to enhance the usability of language processing tools.
Improving LLMs’ Learning for Coreference Resolution
This paper introduces novel techniques to enhance Coreference Resolution in LLMs, addressing issues of hallucination through improved training methodologies. The integration of these methods shows significant promise for advancing the capabilities of language models in this domain.