The author addresses the challenge of semantic hallucinations in NLG models by proposing an automatic pipeline for detection, utilizing data augmentation and an ensemble of methodologies. The main thesis is to improve the accuracy of identifying hallucinations in generated text.
大規模言語モデルを使用した幻覚検出の自動パイプラインを提案し、SemEval-Task 6 SHROOMで80.07%の精度を達成。