News Recommendation¶
Code for comparing news article representations with PLM-NR.

Read the online documentation for the user guide and API reference.
See SETUP.md for installation, data downloads, model paths, and parameter examples.
The experiments compare three article representations using the same ranking architecture, training settings, and chronological data split.
| Representation | Article input (512 BERT tokens) |
|---|---|
title_abstract_512 |
Title and original abstract |
title_section_summary_512 |
Title and summaries of predicted topic segments |
title_article_summary_512 |
Title and a generated whole-article summary |
Quick start¶
Follow installation and supply the data and weights, then rebuild the saved experiment inputs from the repository root:
cd recommendation
source scripts/activate.sh
python -m news_recommendation.pipeline replay --output runs/replay
This step runs on CPU. replay rebuilds the split and packs the saved segmentation, summaries, and segment selection into runs/replay/data/.
Reproduce the thesis experiments¶
Run from recommendation/ with the environment activated. Train and evaluate all three representations using the inputs prepared above:
python scripts/run_training.py --data runs/replay/data --gpus 0 1 2 3 --output-dir runs/reproduction
The launcher trains each representation separately and evaluates the epoch with the best validation AUC. Runs save model weights, configuration, validation metrics in epochs.json, and test AUC, MRR, nDCG@5, and nDCG@10 in test.json. Existing results are protected from overwriting.
To inspect the segment-summary results:
cat runs/reproduction/title_section_summary_512/test.json
Training settings are in configs/plm_nr.json. See run commands to select one representation or regenerate segmentation and summaries.
Folder structure¶
recommendation/
├── configs/ # experiment settings, paths, and summary prompts
├── scripts/ # environment and training commands
├── src/news_recommendation/ # data preparation, summaries, ranking, and scoring
├── generation/ # summary-generation environment
├── data/ # MIND articles, impressions, and prepared inputs
├── checkpoints/ # downloaded model weights
└── runs/ # trained models, predictions, and scores
Third-party code¶
The PLM-NR model is adapted from wuch15/PLM4NewsRec by Chuhan Wu and colleagues.
Citations¶
@inproceedings{DBLP:conf/sigir/WuWQ021,
author = {Chuhan Wu and
Fangzhao Wu and
Tao Qi and
Yongfeng Huang},
title = {Empowering News Recommendation with Pre-trained Language Models},
booktitle = {Proceedings of the 44th International ACM SIGIR Conference
on Research and Development in Information Retrieval},
pages = {1652--1656},
publisher = {ACM},
year = {2021},
url = {https://doi.org/10.1145/3404835.3463069},
doi = {10.1145/3404835.3463069}
}
@inproceedings{DBLP:conf/acl/WuQCWQLLXGWZ20,
author = {Fangzhao Wu and
Ying Qiao and
Jiun{-}Hung Chen and
Chuhan Wu and
Tao Qi and
Jianxun Lian and
Danyang Liu and
Xing Xie and
Jianfeng Gao and
Winnie Wu and
Ming Zhou},
title = {{MIND:} {A} Large-scale Dataset for News Recommendation},
booktitle = {Proceedings of the 58th Annual Meeting of the Association for Computational
Linguistics, {ACL} 2020, Online, July 5-10, 2020},
pages = {3597--3606},
publisher = {Association for Computational Linguistics},
year = {2020},
url = {https://doi.org/10.18653/v1/2020.acl-main.331},
doi = {10.18653/V1/2020.ACL-MAIN.331}
}