🧠 Cogito Ergo Summ

"I think, therefore I summarize"

Overview

Code and data accompanying the AAAI23 paper "Cogito Ergo Summ: Abstractive Summarization of Biomedical Papers via Semantic Parsing Graphs and Consistency Rewards".

CogitoErgoSumm is the first framework for single-document biomedical abstractive summarization equipping large pre-trained language models with rich domain-specific and domain-general semantic parsing graphs: events and AMRs. Event and AMR graph embeddings are learned by edge-aware graph attention networks. We propose new decoder cross-attention modules, and design a reinforcement learning (RL) reward signal to preserve source-summary semantics consistency.

Experiments and ablation studies on CDSR demonstrate that our framework sets new marks in informativeness, factuality, and readability, better selecting and preserving summary-worth content.

🔎 Paper

Read our paper

🔎 Poster

Read our poster

🔎 Supplementary Material

Read our supplementary material

✉ Contacts

• Giacomo Frisoni, giacomo.frisoni[at]unibo.it
• Paolo Italiani, paolo.italiani[at]studio.unibo.it
• Stefano Salvatori, s.salvatori[at]unibo.it
• Gianluca Moro, gianluca.moro[at]unibo.it

If you have troubles, suggestions, or ideas, the Discussion board might have some relevant information. If not, you can post your questions there 💬🗨.

License

This project is released under the CC-BY-NC-SA 4.0 license (see LICENSE).

Cite

If you use CogitoErgoSumm in your research, please cite:

@article{frisoni-etal-2023-cogitoergosumm,
  title     = {Cogito Ergo Summ: Abstractive Summarization of Biomedical Papers via Semantic Parsing Graphs and Consistency Rewards},
  author    = {Giacomo, Frisoni and Paolo, Italiani and Gianluca, Moro and Stefano, Salvatori},
  booktitle = {Thirty-Seventh {AAAI} Conference on Artificial Intelligence, {AAAI} 2023},
  pages     = {1--10},
  publisher = {{AAAI} Press},
  year      = {2023}
}