CO2Sum

• Code for the paper CO2Sum:Contrastive Learning for Factual-Consistent Abstractive Summarization
• Include the negative sample constuction method LFN CO2Sum_LFN, and the training code CO2Sum_train

CO2Sum_LFN

• Code for the LFN algorithm introduced in the paper
• The process of LFN consists of three steps:
  • run get_next.py to get the context of summary
  • run ./LFN_LM/run.sh to get the fact fragments based on the article, summary and context
  • run LFN_construct.py to construct the negative samples based on the fact fragments

CO2Sum_train

• We develop our method based on the fairseq. Since there is no model architecture modified, you can just extend the criterion, data, tasks by setting --user-dir to CO2Sum_train then start training and inference by using default fairseq-train and fairseq-generate
• The loss function of CoEnc and CoDec are described in the ./criterions/label_smoothed_cross_entropy_with_position_triplet_contrastive.py
• The data loading process for ground truth summary and negative samples is described in ./data/language_position_triplet_dataset.py