Submitted to TACL
This paper introduces MURI, a novel method for generating high-quality instruction tuning datasets for low-resource languages without human annotators.
Submitted to TACL
This paper proposes CRAFT, a method for generating synthetic datasets for specialized tasks using a few user-written examples.
EMNLP 2024 Findings
This paper introduces SYNTHEVAL, a hybrid approach for evaluating NLP models through dynamic behavioral testing.
EMNLP 2024 Findings
This paper presents TurkishMMLU, the first multitask, multiple-choice QA benchmark designed specifically for Turkish.
EMNLP 2024 Findings
This paper presents CovEReD, a counterfactual data generation approach for document-level relation extraction datasets.
arXiv
This paper introduces MemLLM, a method for finetuning LLMs to use an explicit read-write memory, addressing limitations in knowledge-intensive tasks.
arXiv
This paper demonstrates the usefulness of Construction Grammar through Argument Structure Constructions and proposes a hybrid human-LLM approach for corpus construction.
EMNLP 2024 Findings
This paper presents LongForm-C, a dataset created through reverse instructions to enhance instruction tuning for language models.
EMNLP 2023 Findings
This paper introduces MEAL, a method for stable and active learning in few-shot prompting to address high variance in data selection and run variability.
EMNLP 2023 Findings
This paper introduces LABDet, a robust and language-agnostic method for evaluating social bias in pretrained language models.
arXiv
This paper proposes a novel method using counterfactual open book QA to improve attribution in language models.
arXiv
This paper introduces HATELEXICON, a lexicon of slurs and targets of hate speech for multiple countries, and demonstrates its use in interpreting model predictions.