Researchers reveal major challenges in applying standard AI sentiment analysis to Holocaust oral histories, finding significant disagreement among leading language models.

Study evaluated three pretrained transformer-based sentiment classifiers on 107,305 utterances from Holocaust oral history corpus, revealing substantial performance degradation under domain shift

Introduced ABC (agreement-based stability) taxonomy to categorize inter-model disagreement and identify systematic failure patterns across 579,013 sentences

Employed supplementary T5-based emotion classifier to analyze emotional distributions across agreement strata, providing deeper insights into model behavior variations

Research highlights that complex, long-form narratives with intricate discourse structures pose significant challenges for off-the-shelf sentiment detection systems