Towards Automated Classification of Adult Attachment Interviews in German Language Using the BERT Language Model

Attachment theory, pioneered by John Bowlby, has become a prominent psychological framework that aids in our comprehension of human behavior within close relationships. The Adult Attachment Interview (AAI), a semi-structured interview, provides a standardized method for assessing adult attachment styles, offering insights into an individual’s attachment patterns, emotional regulation, and relational experiences. However, manual AAI classification is a labour-intensive and highly specialized task. Thus, automating this process can optimize the classification of individual attachment patterns and their psychological and relational implications, thereby increasing efficiency and accuracy in assessing attachment styles. In this work, we investigate the application of a BERT large language model and linguistic features for the automated classification of transcribed Adult Attachment Interviews conducted in German into the three categories: secure, dismissing, and preoccupied. The findings of our study indicate that using BERT embeddings alone can yield results comparable to, and in some cases better than, those achieved with traditional linguistic features. However, this effect should be interpreted with caution, as it does not hold consistently across all settings. Another key conclusion of our study is that augmenting the original AAI dataset with artificially generated interviews produced by state-of-the-art large language models generally improves predictive performance across a range of classification models.