Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Combining Hierachical VAEs with LLMs for clinically meaningful timeline summarisation in social media

Jiayu Song, Jenny Chim, Adam Tsakalidis, Julia Ive, Dana Atzil-Slonim, Maria Liakata

TL;DR

This work proposes Timeline Hierarchical VAE (TH-VAE), a hybrid abstractive method that combines a hierarchical variational autoencoder with prompting-enabled LLMs to produce clinically meaningful, temporally sensitive summaries from social media timelines. TH-VAE generates first-person evidence timelines, while an instruction-tuned LLM creates a separate high-level clinical summary in the third person; the two outputs are designed to support mental health monitoring and clinical decision making. Evaluation on expert-annotated data and human judgments shows TH-VAE provides superior temporal coherence and factuality compared to LLM-only approaches, with the added benefit of clinically grounded prompts enhancing the usefulness of the high-level summaries. The approach demonstrates a practical pathway to automated, clinically aligned longitudinal summaries from user-generated social data, while acknowledging ethical and reliability considerations for real-world deployment.

Abstract

We introduce a hybrid abstractive summarisation approach combining hierarchical VAE with LLMs (LlaMA-2) to produce clinically meaningful summaries from social media user timelines, appropriate for mental health monitoring. The summaries combine two different narrative points of view: clinical insights in third person useful for a clinician are generated by feeding into an LLM specialised clinical prompts, and importantly, a temporally sensitive abstractive summary of the user's timeline in first person, generated by a novel hierarchical variational autoencoder, TH-VAE. We assess the generated summaries via automatic evaluation against expert summaries and via human evaluation with clinical experts, showing that timeline summarisation by TH-VAE results in more factual and logically coherent summaries rich in clinical utility and superior to LLM-only approaches in capturing changes over time.

Combining Hierachical VAEs with LLMs for clinically meaningful timeline summarisation in social media

TL;DR

This work proposes Timeline Hierarchical VAE (TH-VAE), a hybrid abstractive method that combines a hierarchical variational autoencoder with prompting-enabled LLMs to produce clinically meaningful, temporally sensitive summaries from social media timelines. TH-VAE generates first-person evidence timelines, while an instruction-tuned LLM creates a separate high-level clinical summary in the third person; the two outputs are designed to support mental health monitoring and clinical decision making. Evaluation on expert-annotated data and human judgments shows TH-VAE provides superior temporal coherence and factuality compared to LLM-only approaches, with the added benefit of clinically grounded prompts enhancing the usefulness of the high-level summaries. The approach demonstrates a practical pathway to automated, clinically aligned longitudinal summaries from user-generated social data, while acknowledging ethical and reliability considerations for real-world deployment.

Abstract

We introduce a hybrid abstractive summarisation approach combining hierarchical VAE with LLMs (LlaMA-2) to produce clinically meaningful summaries from social media user timelines, appropriate for mental health monitoring. The summaries combine two different narrative points of view: clinical insights in third person useful for a clinician are generated by feeding into an LLM specialised clinical prompts, and importantly, a temporally sensitive abstractive summary of the user's timeline in first person, generated by a novel hierarchical variational autoencoder, TH-VAE. We assess the generated summaries via automatic evaluation against expert summaries and via human evaluation with clinical experts, showing that timeline summarisation by TH-VAE results in more factual and logically coherent summaries rich in clinical utility and superior to LLM-only approaches in capturing changes over time.
Paper Structure (31 sections, 5 equations, 3 figures, 6 tables)

This paper contains 31 sections, 5 equations, 3 figures, 6 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Methodology
  4. Architecture Overview
  5. Input to Timeline Summarisation
  6. Overview of TH-VAE
  7. Document Reconstruction via TH-VAE
  8. Hierarchical Component
  9. High-level Mental Health Summarization
  10. Experiments
  11. Evaluation Dataset Creation
  12. Models & Baselines
  13. Evaluation
  14. Results
  15. Automatic evaluation
  16. ...and 16 more sections

Figures (3)

  • Figure 1: Prompting framework for generating high-level summaries. Taking a first-person summarised timeline as input, we (1) prompt the LLM around different clinical topics, (2) summarise extracted inferences into prose per topic, and (3) combine the topic-specific intermediate summaries into a coherent, distilled document.
  • Figure 2: Each timeline is separated into several segments based on 'MoC'. We highlight the key phrases.
  • Figure 3: Overview of TH-VAE. The left of the dotted line shows the construction of the k-sentence representation used only during generation, informed by the key-phrases, while the right side shows the hierarchical structure of TH-VAE, and its components.① and ② represent the input during training and generation respectively.