Training Models to Generate, Recognize, and Reframe Unhelpful Thoughts

1 Introduction

• Self-help CBT has shown promise, but insufficient practice material suited to people’s specific circumstances is a barrier to adoption.CBT includes identifying and reframing unhelpful ways of thinking.Source: §1 · p. 1
• Existing public reframing data use stress-marked tweets and do not target CBT thought-pattern categories or examples tailored to particular situations.Much other NLP work on CBT uses patient-professional interactions that are not publicly available.Source: §1 · p. 1
• PATTERNREFRAME contains about 10k persona-conditioned thoughts exhibiting ten classical unhelpful thought patterns, paired with crowdsourced reframes.The dataset supports generating and reframing thoughts given a persona and pattern, as well as classifying a thought’s pattern.Source: §1 · p. 1

2.1 NLP for Mental Health

• Earlier unhelpful-thought datasets were built from patient-therapist interactions and are not publicly available.Other mental-health NLP studies commonly annotate social-media posts and train classification models.Source: §2.1 · p. 2
• Ziems et al. (2022) introduced reframing for stress-marked tweets, whereas this work adds persona and classical CBT-pattern conditioning and extends the tasks to generating and identifying thoughts.The prior source material was neither persona-conditioned nor focused on classical unhelpful thought patterns.Source: §2.1 · p. 2

2.2 Controllable Text Generation

• Controllable text generation methods include prompting, fine-tuning, post-processing, and pretraining; this work experiments with prompting and fine-tuning.The passage describes control tokens and task-specific reward functions among the approaches.Source: §2.2 · p. 2

3 Identifying and Reframing Unhelpful Thoughts

• The ten unhelpful-thought categories are described in lay terms in a CBT self-help book used for bibliotherapy.The categories are derived from Burns (1980).Source: §3 · p. 2
• The five proposed reframing strategies are Growth Mindset, Optimism, Impermanence, Neutralizing, and Self-affirmation.They respectively emphasize learning, positive aspects, accepting temporary adversity, realistic alternatives, and core values.Source: §3 · p. 2
• Workers may use reframing strategies beyond the five suggestions, including mindfulness or forgiveness and compassion.The listed strategies are a starting point, not a restriction on workers’ choices.Source: §3 · p. 2

4.1.1 Task 1: Writing Unhelpful Thoughts

• Models trained or prompted with the dataset are evaluated on generating persona- and pattern-conditioned thoughts and reframes, and classifying patterns.The collection begins with a persona and one of ten patterns as context for writing a matching thought.Source: §4.1.1 · p. 2, §4.1 · p. 2

4.1.2 Task 2: Categorizing Unhelpful Thoughts

• Workers label all patterns present in each thought, reflecting that unhelpful thoughts can exhibit overlapping patterns.Five annotations are collected per thought, and thoughts marked “None” by a majority are discarded.Source: §4.1.2 · pp. 2–3

4.1.3 Task 3: Reframing Unhelpful Thoughts

• Workers produce three context-aligned rewrites per thought, aiming to remove its unhelpful pattern while selecting a strategy if applicable.The five suggested strategies are optional; workers may use other appropriate approaches.Source: §4.1.3 · p. 3

4.1.4 Task 4: Evaluating the Rewrites of Unhelpful Thoughts

• Rewrites are retained only when a majority of annotators judge them valid: removing the pattern without logically contradicting the source.Five annotations are collected for each set of rewrites.Source: §4.1.4 · p. 4

4.2 Data Quality

• The collection process qualified 524 workers, with selected workers used for evaluation and their annotations included in the final dataset.Onboarding tasks checked workers’ understanding of reframing; agreement was 0.355 for task two and 0.454 for task four.Source: §4.2 · p. 4

4.3.1 Dataset Statistics

• 9,688 thoughts and 26,507 reframed versions make up PATTERNREFRAME, averaging 2.74 retained rewrites per thought.Thoughts averaged 19.1 words and rewrites 23.9 words.Source: §4.3.1 · p. 4

4.3.2 Analysis of Unhelpful Thought Patterns

• Pattern labels overlap substantially for several related categories, while All or Nothing Thinking overlaps with many categories and Mind Reading and Should Statement overlap least.Figure 1 compares the pattern used to collect each thought with annotators’ selected patterns.Source: §4.3.2 · p. 4, Figure 1 · p. 4

4.3.3 Analysis of Reframing Strategies:

• Growth Mindset and Optimism are the most common strategies, and Optimism is especially common for Discounting the positive and Mental Filtering.Most rewrites use one of the five strategies, with few labeled “None.”Source: §4.3.3 · p. 5, Figure 2 · p. 5

5.1.1 Task and Data

• Generation conditions on a persona and thought pattern; models include fine-tuned BART-large, T5-large, and R2C2-3B, plus one-shot GPT3.5, with BLEU, ROUGE, BERTScore, and diversity metrics.GPT3.5 generation evaluation used 100 random test inputs because API access was limited.Source: §5.1.1 · p. 5, §5.1.2 · p. 5, §5.1.3 · p. 5

5.1.3 Automatic Evaluation

• For generating thoughts, models perform similarly on BLEU, BERTScore, and ROUGE, while GPT3.5 achieves the best Distinct-n scores.The reported automatic metrics measure similarity to human references and lexical diversity.Source: §5.1.3 · p. 5, Table 2 · p. 6

5.1.4 Human Evaluation

• In generation human ratings, GPT3.5 outperforms other approaches and human references on fluency and pattern quality, but has the lowest meaning-preservation score, which remains very high.Outputs count as positively rated when at least five of nine annotators agree.Source: §5.1.4 · p. 6, Figure 3 · p. 6

5.2.1 Task and Data

• Classification assigns a persona-thought pair to one of ten patterns or None; RoBERTa uses soft labels to reflect pattern overlap, while GPT3.5 uses one example per category.The augmented dataset includes negative examples representing neutral texts and mismatched thoughts and personas.Source: §5.2.1 · p. 6, §5.2.2 · pp. 6–7

5.2.3 Evaluation

• RoBERTa performs above 72% on all clustered categories except Mislabeling, while one-shot GPT3.5 performs worse and shows prompt-order sensitivity.The None category performs best; 40% of GPT3.5 classification predictions changed after prompt examples were reordered.Source: §5.2.3 · p. 7, Figure 4 · p. 7

5.3.1 Task and Methods

• Reframing takes the persona, pattern, and unhelpful thought as input and aims to produce a persona-consistent thought without that pattern, using the generation task’s evaluated approaches.Fine-tuned methods concatenate the three inputs; few-shot methods use prompts with reframed examples.Source: §5.3.1 · p. 7

5.3.2 Automatic Evaluation

• For reframing, R2C2 performs best on BLEU and BERTScore, while GPT3.5 outperforms other models and human references on Distinct-1 and Distinct-2.Automatic scores use the maximum across up to three human reframes for each thought, then average those maxima.Source: Table 2 · p. 6, §5.3.2 · pp. 7–8

5.3.3 Human Evaluation

• All models exceed 70% on reframing quality, with GPT3.5 best; R2C2 and GPT3.5 are fluent and preserve meaning, and GPT3.5 performs close to crowdsource workers using prompting alone.The more stringent threshold of seven positive ratings out of nine makes GPT3.5’s superiority more marked.Source: Figure 3 · p. 6, §5.3.3 · p. 8

6 Conclusion

• PATTERNREFRAME contains about 10k persona-conditioned unhelpful thoughts and multiple complementary reframes for each.The reframes avoid the initial unhelpful pattern and express the thought more constructively.Source: §6 · p. 8
• Modern language models can generate, identify, and reframe unhelpful thoughts conditioned on a persona.The authors suggest the dataset could provide practitioners with richer, more diverse examples addressing a lack of personalized and specific examples for teaching cognitive techniques.Source: §6 · pp. 8–9
• Future work will test whether richer model-generated training material improves human learning and understanding of unhelpful thought patterns.The authors propose this work may support psychological validation studies and future low-intensity self-help interventions.Source: §6 · p. 9

7 Limitations

• Persona sources are English-language responses from US workers, so generated examples may not reflect thought patterns and personas across cultures and diverse populations.The data was also generated by paid workers rather than people genuinely engaging with situations that matter to them; models generate only English.Source: §7 · p. 9
• The collected material reflects lay understanding, and the full dataset still requires further validation despite a generally sound psychologist spot-check.The spot-check was conducted by a CBT-trained clinical psychologist.Source: §7 · p. 9
• The tested models are resource-intensive, and the best-performing model, GPT3.5, is available only through a paid API.The resource demands constrain access to the tested systems.Source: §7 · p. 9

8 Ethical considerations

• The paired reframing data could be used in reverse to generate negative thoughts from positive or neutral text.The authors relate this risk to toxicity reduction and sentiment style-transfer tasks.Source: §8 · p. 9
• Unsupervised use could be harmful because the material has not been tested with participants while monitoring adverse events.The authors name increased anxiety and distorted understanding of unhelpful thoughts or useful reframings as possible adverse events.Source: §8 · p. 9

A.1 Onboarding Tasks

• Workers had to identify unhelpful thoughts and suitable reframes in onboarding, passing three of four checks before participating.Separate onboarding tasks covered the first two data-collection tasks and the later reframing and evaluation tasks.Source: §A.1 · p. 13

B Data Collection Interface Snapshots

• The collection interfaces guide workers to write unhelpful thoughts, select their patterns, reframe them, and evaluate reframing quality.These are the first four tasks in the data-collection pipeline.Source: Figure 5 · p. 13, Figure 6 · p. 13, Figure 7 · p. 14

C Evaluation Interface Snapshots

• Evaluation interfaces are provided for generated thoughts and reframed thoughts.Figures 9 and 10 show the two evaluation interfaces.Source: Figure 9 · p. 15, Figure 10 · p. 15

D.1 Generation Models

• Generation baselines BART, T5, and R2C2 were fine-tuned with validation-based checkpoint selection and beam search at inference.The models used beam size 10; checkpoint selection was based on validation perplexity.Source: §D.1 · p. 16

D.2 Classification Models

• Classification experiments fine-tuned RoBERTa-large for up to 10 epochs, selecting checkpoints by validation weighted F1.The setup used a learning rate of 1e-05 and batch size 16.Source: §D.2 · p. 16

E GPT3.5 Prompt Examples

• The appendix provides example GPT3.5 prompts for generating, reframing, and identifying unhelpful thoughts.Source: Table 4 · p. 16, Table 5 · p. 16, Table 6 · p. 17

• Figure 11 compares generation and reframing outputs by the percentage rated positively by at least seven of nine annotators.Source: Figure 11 · p. 18