Training Models to Generate, Recognize, and Reframe Unhelpful Thoughts

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 13641–13660 July 9-14, 2023 ©2023 Association for Computational Linguistics Training Models to Generate, Recognize, and Reframe Unhelpful Thoughts Mounica Maddela∗ Georgia Tech & Meta AI Megan Ung∗ Meta AI Jing Xu Meta AI Andrea Madotto Meta AI Heather Foran Klagenfurt University Y-Lan Boureau Meta AI Abstract Many cognitive approaches to well-being, such as recognizing and reframing unhelpful thoughts, have received considerable empiri- cal support over the past decades, yet still lack truly widespread adoption in self-help format. A barrier to that adoption is a lack of adequately specific and diverse dedicated practice material. This work examines whether current language models can be leveraged to both produce a vir- tually unlimited quantity of practice material illustrating standard unhelpful thought patterns matching specific given contexts, and generate suitable positive reframing proposals. We pro- pose PATTERNREFRAME, a novel dataset of about 10k examples of thoughts containing un- helpful thought patterns conditioned on a given persona, accompanied by about 27k positive reframes. By using this dataset to train and/or evaluate current models, we show that exist- ing models can already be powerful tools to help generate an abundance of tailored practice material and hypotheses, with no or minimal additional model training required. 1 Introduction Cognitive Behavioral Therapy (CBT) (Beck, 1963, 1976) is one of the most robustly validated ap- proaches in psychology (Hofmann et al., 2012; David et al., 2018). A core pillar of CBT consists in identifying and reframing unhelpful ways of think- ing. Low-intensity CBT interventions have shown promise in self-help formats (Shafran et al., 2021; Williams, 2001), yet a lack of sufficient practice material suited to people’s specific circumstances is a barrier to adoption (Helgadóttir et al., 2009). Through prompting, control tokens, or adequate conditioning, modern language models can guide generation of language towards desired outcomes, such as conforming to a given persona (Zhang et al., 2018), style (Ziems et al., 2022), or level of con- fidence (Mielke et al., 2022). This makes them a potentially powerful practice aid for learning cog- nitive reframing techniques. A major barrier is the lack of publicly available data. Most exist- ing work in natural language processing (NLP) for CBT focuses on interactions between patients and mental health professionals, which are not pub- licly available (Mieskes and Stiegelmayr, 2018; Rojas-Barahona et al., 2018; Shreevastava and Foltz, 2021). Ziems et al. (2022) released the first public dataset for reframing tweets marked with a hashtag indicating stress, using known reframing techniques, but it does not specifically look at the categories of unhelpful thinking used in CBT, and uses existing tweets rather than allowing the gener- ation of examples suited to a particular situation. In this work, we propose1 a novel dataset, PAT- TERNREFRAME, consisting in ∼10k crowdsourced examples of thoughts containing ten classical types of unhelpful thought patterns (Burns, 1980), con- ditioned on personas, matched with crowdsourced proposals of reframing that do not exhibit the pat- terns. We introduce two …

Content truncated to 500 words