This project focuses on developing an NLP-based system for generating concise and informative summaries of mental health counselling conversations. The aim is to capture the key insights of therapeutic dialogues while preserving their contextual nuances.
The MEMO dataset used in this project consists of therapist-client conversations, providing a rich source of dialogue data for summarization tasks.
The first method involved a two-stage processing of dataset. We use a pre-trained model to tag each utterance in the conversation with an appropriate emotion. Then we finetuned T5/Pegasus on emotion-annotated data.
The second method was multi-dataset finetuning approach. We fintetuned T5 on GoEmotions and DialogSum datasets so the model learns about the emotional context and dialog summarization task. Then we finetune this model again on the MEMO dataset. All finetuning here is done with LoRA.
More details about the baselines and implementation can be found in report.pdf.
Related Research Paper: Counseling Summarization Using Mental Health Knowledge Guided Utterance Filtering