Code for Yeyang Liu's master's thesis at the University of Zurich: news topic segmentation from editorial subheadings, clinical transfer, and applications to information extraction and recommendation.
Read the online documentation or build the three documentation sites locally.
| # | Research question | Part |
|---|---|---|
| RQ1 | How can editorial structure be used to train effective news topic segmentation models, and how well do these models transfer to clinical documents? | Topic segmentation |
| RQ2 | How does input segmentation affect information extraction for knowledge graph construction in news and clinical documents? | Knowledge graph |
| RQ3 | How does representing news articles through summaries of predicted topic segments affect recommendation ranking, compared with abstract and whole-article representations? | Recommendation |
Each part has its own src/, scripts/, and README. Segmentation and recommendation use configs/; knowledge-graph settings live beside the model adapters.
Install uv, enter the part you want to use, and activate its locked environment:
cd topic_segmentation # or knowledge_graph / recommendation
source scripts/activate.shFollow the part's README for experiment commands and its setup guide for data and model assets: segmentation, knowledge graph, recommendation. Recommendation summary generation uses a separate environment from ranking.
Datasets, model weights, and run outputs are not in git.
- MIMIC-IV-Note requires credentialed PhysioNet access. No MIMIC text, MIMIC-derived prediction, or MIMIC-trained model is distributed with this repository.
- MIND, C4, and DWIE must be obtained from their original sources.
The segmentation code builds on Yu et al. (2023); see its citations. Other sources are listed under knowledge-graph third-party code and recommendation third-party code.
See third-party notices for the applicable licenses and components whose upstream repositories do not state a license.