TRILOGUE introduces a benchmark for spoken dialogue fact-checking in three languages, addressing challenges posed by ASR errors and cross-lingual transfer. It contains nearly 12,000 dialogues, 187,000 turns, and 390 hours of paired audio with transcripts and timestamp alignments.
The dataset supports claim check-worthiness detection, source evidence retrieval, and claim verification using various input types, including claim-only, gold evidence, and retrieved evidence. It includes human-recorded dialogue files in Russian and Kazakh.
Baselines indicate that ASR degradation and cross-lingual transfer remain difficult, especially for Kazakh. Retrieved source evidence helps improve verification accuracy compared to using only gold evidence.
This resource provides a foundation for developing and evaluating end-to-end spoken dialogue fact-checking systems across multiple languages.
Source: https://arxiv.org/abs/2609.04452