AutoMin2023

Datasets

Minuting datasets

• ELITR-minuting-corpus (small)
• Additional data generated by ChatGPT from transcripts?

Relevant summarization datasets

• HuggingFace: https://huggingface.co/datasets?task_categories=task_categories:summarization

• CNN-Daily Mail

  • on HuggingFace
  • CNN articles and summaries

• AMI Meeting Corpus

  • on HuggingFace
  • product meetings transcripts and summaries

• ICSI Meeting Corpus

  • not on HuggingFace
  • academic meetings transcripts and summaries
  • code to process: https://github.com/xcfcode/meeting_summarization_dataset

• Spotify Podcast Dataset

  • not on HuggingFace

• SAMSum Corpus

  • on HuggingFace
  • messenger-like conversations with summaries

• DialogSum

  • on HuggingFace
  • dialogues and summaries

• XSum Dataset

  • on HuggingFace
  • news articles and summaries

• MediaSum

  • on HuggingFace
  • media interview transcripts and summaries

• OAGK/OAGKX

  • on Lindat
  • scientific articles with abstracts

• QMSum

  • not on HuggingFace
  • meeting transcripts, query-based summarizations of various topics discussed

• Info:

  • AMI and ICSI best but rather small

ELITR Minuting Corpus

Minutes

• Original
  • Taken real time, some content of the meeting might be missing
• Generated
  • Taken later by independent annotator not present in the meeting
• Both can be used, both have some problems

Alignments

• They are aligned manually
• Can also be used for training, extract smaller pieces of transcripts and minutes that are properly aligned
• We can talk to Marie Hledikova, email or Wednesday at 10:00

Models

Pretrained summarization models

• https://huggingface.co/models?pipeline_tag=summarization&sort=downloads

• Relevant:

  • https://huggingface.co/vmarklynn/bart-large-cnn-icsi-ami-v3
  • https://huggingface.co/philschmid/flan-t5-base-samsum
  • https://huggingface.co/lidiya/bart-large-xsum-samsum

• DialogLM (model for long dialogue understanding) https://github.com/microsoft/DialogLM

Implementations

Minuting Baseline Experiments

• Many work done, but not documented properly
• Some documentation: https://elitr.eu/deliverables/

Additional Information related to the topic

• https://nlp-yang.github.io/Dialogue_Summarization.pdf
• https://www.catalyzex.com/s/Meeting%20Summarization

Computing

Running jupyter notebook on UFAL servers

• https://nero-docs.stanford.edu/jupyter-slurm.html
• https://stackoverflow.com/questions/35545402/how-to-run-an-ipynb-jupyter-notebook-from-terminal

TODO

1. Try some HuggingFace summarization pretrained models and finetune it on the minuting dataset
2. Download and analyze the various datasets
3. Preprocess the Europarl dataset