Generate model finetune:
t5 training for writing poems and couplets: https://github.com/hululuzhu/chinese-ai-writing-share
Fengshenbang-LM version
t5 installation: https://github.com/IDEA-CCNL/Fengshenbang-LM/blob/main/fengshen/examples/qa_t5/finetune_t5_cmrc.py
t5 installation summary: https://github.com/IDEA -CCNL/Fengshenbang-LM/blob/main/fengshen/examples/mt5_summary/pretrain_mt5_summary.shReference5
: https://github.com/IDEA-CCNL/Fengshenbang-LM/tree/main/fengshen/examples/pretrain_t5
textgen framework
https://github.com/shibing624/textgen
UDA (non-core word replacement)/EDA: This project refers to Google's UDA (non-core word replacement) algorithm and EDA algorithm, and based on TF-IDF, the part of the sentence is not important Words are replaced by synonyms, random words are inserted, deleted, replaced and other methods to generate new text and realize text amplification.
BT (Back Translation): This project implements the back translation function based on Baidu translation API, and first translates Chinese sentences into English , and then translate English into new Chinese
Seq2Seq: This project implements the training and prediction of Seq2Seq, ConvSeq2Seq, and BART models based on PyTorch, which can be used for text translation, dialogue generation, summary generation and other text generation tasks. T5: This project is implemented based on
PyTorch T5 and CopyT5 model training and prediction can be used for text generation tasks such as text translation, dialogue generation, couplet generation, and copywriting. GPT2:
This project implements GTP2 model training and prediction based on PyTorch, which can be used for article generation, couplet generation, etc. Text generation task
SongNet: This project implements SongNet model training and prediction based on PyTorch, which can be used for text generation tasks such as poems and lyrics in standardized formats. TGLS: This project
implements the TGLS unsupervised similar text generation model, which is a "first search Post-learning" text generation method, through iteratively learning the candidate set, the final model can generate high-quality similar text similar to the candidate set