This repo contains the code for the Findings of EMNLP 2022 paper "He Said, She Said: Style Transfer for Shifting the Perspective of Dialogues"
for reproducing extractive summarization results: refer to PreSumm for their implementation of their extractive summarization model.
All models were trained on either a v100 32GB GPU or a 2080TI GPU, with batch sizes/gradient accumulation steps adjusted for the differing memory size to maintain the same effective batch size.
On a 2080TI GPU, all BART-large perspective shifting models trained in under 6 hours. On a v100 GPU, these models trained in less than 2 hours. Extractive summarization experiments were run on a 2080TI GPU and trained in less than 30 minutes when early stopping was used with a batch size of 1500.
The BART-large model contains approximately 406M parameters. The BERT-base-uncased models used for extractive summarization contain approximately 120M parameters.