Speech frustration responses: an automatic categorization.

Abstract

Our study aimed to construct an automatic classification tool for verbal responses to the Rosenzweig test with participation of a psychologist, a linguist and an artificial intelligence specialist. We describe verbal constructions specific to four out of nine types of frustration responses. For each type of response, we provide data on the effectiveness of the proposed tool for automatic classification (completeness and reliability). The study pipeline was as follows: first, 462 Rosenzweig test protocols were processed by a psychologist and made up a research corpus that was further analyzed by a linguist; next, the linguistic descriptions were applied in order to build a feature description of text fragments; finally, the classifier of frustration reaction types derived from the feature descriptions. The classification accuracy was evaluated by means of statistical procedure of three-fold cross-sliding control. Ourlinguistic patterns form a high-level feature description of text fragments, which allows identifying the type of reaction statements with high completeness (R not less than 0.8). Four responses to frustration -M, M’, I, E -were reliably distinguished beyond the context of statements (F1 > 0.7). The results of linguistic analysis of the verbal reactions to the Rosenzweig test related to the four types of frustration responses are discussed;4 to 18 linguistic descriptions were obtained for different types of responses. Within psychology the described technology of linguistic patterns represents a means for professional reflection, a tool for the verification of projective methods results. It also contributes to the improvement of automatic speech diagnostics (e.g., for online discussions).