%Aigaion2 BibTeX export from Idiap Publications
%Friday 18 September 2026 04:31:06 PM

@INPROCEEDINGS{Watawana_LREC2026_2026,
                      author = {Watawana, Hasindri Sankalpana and Burdisso, Sergio and Moreno-Galvan, Diego Aaron and S{\'{a}}nchez-Vega, Fernando and L{\'{o}}pez-Monroy, A. Pastor and Motlicek, Petr and Villatoro-Tello, Esa{\'{u}}},
                    projects = {Idiap, ORIENTER},
         mainresearchprogram = {AI for Life},
  additionalresearchprograms = {AI for Everyone},
                       month = may,
                       title = {When Consistency Becomes Bias: Interviewer Effects in Semi-Structured Clinical Interviews},
                   booktitle = {LREC 2026},
                      volume = {Proceedings of the Fifteenth Language Resources and Evaluation Conference},
                        year = {2026},
                   publisher = {ELRA Language Resource Association},
                    location = {Palma de Mallorca, Spain},
                         url = {https://aclanthology.org/2026.lrec-1.185/},
                         doi = {10.63317/34hw23mzd8c7},
                    abstract = {Automatic depression detection from doctor–patient conversations has gained momentum thanks to the availability of public corpora and advances in language modeling. However, interpretability remains limited: strong performance is often reported without revealing what drives predictions. We analyze three datasets—ANDROIDS, DAIC-WOZ, and E-DAIC—and identify a systematic bias from interviewer prompts in semi-structured interviews. Models trained on interviewer turns exploit fixed prompts and positions to distinguish depressed from control subjects, often achieving high classification scores without using participant language. Restricting models to participant utterances distributes decision evidence more broadly and reflects genuine linguistic cues. While semi-structured protocols ensure consistency, including interviewer prompts inflates performance by leveraging script artifacts. Our results highlight a cross-dataset, architecture-agnostic bias and emphasize the need for analyses that localize decision evidence by time and speaker to ensure models learn from participants’ language. Our code is available at: \url{https://github.com/idiap/bias_in_daic-woz/tree/main/LREC_2026}.},
                         pdf = {https://publications.idiap.ch/attachments/papers/2026/Watawana_LREC2026_2026.pdf}
}