Introduction

One implementation of the paper "Dynamic Sliding Window for Meeting Summarization" / "Dynamic Sliding Window Modeling for Abstractive Meeting Summarization" (Interspeech 2022).

Package Requirements

1. pytorch==1.7.1
2. transformers==4.8.2
3. click==7.1.2
4. sentencepiece==0.1.92
5. allennlp==2.6.0
6. allennlp-models==2.6.0

AMI Data Processing

There are 4 steps to process the original AMI meeting conversations, to get the segmented samples described in the paper.

• Step1_build_AMI_data_with_segments.py
  Splitting the meeting conversations as well as the summaries to multiple segments/snippets, this is based on the extractive supporting annotation in AMI data;
• Step2_process_segmented_data.py
  Conducting simple coreference resolution on the summary snippets;
• Step3_build_longABS_target_sentences.py
  The AMI data provides long abstract, to improve the informativeness of generation, we further merge them to the summary snippets;
• Step4_build_AMI_all_utterances.py
  Building the dialogue level sample for testing;

Tranining and Evaluation

The segmented samples are used for training, and the dialogue level samples can be used for evaluation.

• ./AMI_experiments/tmp_data_with_segments/
  There are train/eval/test samples after our pre-processing. For model training, using the train.source (dialogue content) and train.target.plus_long (reference summary);
• ./AMI_experiments/tmp_data_dialogue_level/
  For evaluation for meeting level summarization, using the file train.all.utter.source (meeting level dialogue content) and train.doc_level_longabs (reference summary);

Citation

@inproceedings{Liu2022DynamicSW,
  title={Dynamic Sliding Window Modeling for Abstractive Meeting Summarization},
  author={Zhengyuan Liu and Nancy F. Chen},
  booktitle={Interspeech},
  year={2022},
}

@article{liu2021dynamic,
  title={Dynamic Sliding Window for Meeting Summarization},
  author={Liu, Zhengyuan and Chen, Nancy F},
  journal={arXiv preprint arXiv:2108.13629},
  year={2021}
}