Acoustic Features in Dialogue Dominate Accurate Personality Trait Classification

We report on experiments in identifying personality traits from the dialogue of participants in the MULTISIMO corpus. Experiments used audio and linguistic features from participants’ speech and transcripts, using both self- and observer personality reports. Contrary to our expectations that the linguistic content would best predict traits, the results highlight the multimodal nature of personality computing, suggesting that the content is less important than acoustics: except for two cases, models based on acoustic features only, or combined with linguistic features, outperform models based on linguistic features alone; results also show that there is no optimal choice of a single model or feature set for the prediction of a trait across personality reports, as different models work best for different traits.