Language or rating scales based classifications of emotions: computational analysis of language and alexithymia

Rating scales dominate quantitative assessment of mental health and are often assumed to be more valid than language-based responses. It is unclear how alexithymia (difficulty articulating emotions) affects accuracy of language-based communication of emotions. The study tested whether narratives describing emotional states are more accurately classified by question-based computational analysis of language (QCLA) than by commonly used rating scales, and whether alexithymia moderates accuracy. Phase 1 (N=348) participants generated narratives about depression, anxiety, satisfaction, or harmony. In Phase 2, other participants summarized the narratives using five descriptive words and completed rating scales (PHQ-9, GAD-7, SWLS, HILS). Words were quantified with latent semantic analysis (LSA) and classified via multinomial regression. Language-based responses classified emotional states more accurately than rating scales. Alexithymia did not influence correctness of classification by evaluators for either words or rating scales, but narratives generated by high-alexithymia individuals were harder to classify. Findings suggest language-based responses analyzed computationally may improve assessment of mental health compared with rating scales.

Sverker Sikström, Miriam Nicolai, Josephine Ahrendt, Suvi Nevanlinna, Lotta Stille