Top AI Repos — open-source AI, indexed and scored
Top AI Repos tracks AI repositories on GitHub and answers two different questions about each one: is it moving right now, and would you bet a product on it.
Top AI Repos tracks AI repositories on GitHub and answers two different questions about each one: is it moving right now, and would you bet a product on it.
A Dual Reinforcement Learning Framework for Unsupervised Text Style Transfer (IJCAI 2019)
| Date | Stars |
|---|---|
| 2026-07-31 | 284 |
| 2026-08-01 | 284 |
| 2026-08-02 | 284 |
| 2026-08-06 | 284 |
Today
— stars today
This week
— stars this week
This month
— stars this month
Momentum
0.0
growth rate 0.00%/day
# [A Dual Reinforcement Learning Framework for Unsupervised Text Style Transfer (IJCAI 2019)](https://export.arxiv.org/pdf/1905.10060) <p align="center"><img src="fig/poster.png"></p> ## Reproducibility **In order to help you quickly reproduce the existing works of text style transfer, we release the outputs of all models and the corresponding references.** - **Outputs**: Generated results (outputs) of **10 baselines** and **our model** are in the `outputs/` directory. - **References**: Human references are in the `references/` directory. We also release the three more references we collected on the yelp test dataset, namely `reference[0,1,2,3].0`(the transferred references of negative sentences) and `reference[0,1,2,3].1` (the transferred references of positive sentences). The `reference0.0` and `reference0.1` are collected by [Li et al., 2018](https://github.com/shentianxiao/language-style-transfer). We **strongly recommend** that you use the released multi-references dataset because it has a stronger correlation with human evaluation results. *Ps: We welcome other researchers pull request the outputs of your models.* ## Dataset ### [yelp](https://www.yelp.com/dataset/challenge): negative sentiment (0) <--> positive sentiment (1) - **Original dataset**: The original yelp dataset is in the `data/yelp` directory, where `x.0` denotes the negative `x` type of data and `x.1` denotes the positive `x` type of data. `x` is in `[train, dev, test]`. - **Pseudo-parallel data:** The pseudo-parallel data generated by templates can be found in the `data/yelp/tsf_template` directory. `x.0.tsf` denotes the negative transferred file in which each line only has the sentiment transferred sentence, while `x.0-1.tsf` denotes the negative transferred file in which each line has both the original sentence (input) and sentiment transferred sentence (output). ### [GYAFC](https://github.com/raosudha89/GYAFC-corpus): informal text (0) <--> formal text (1) Since the GYAFC dataset is only free of charge for research purposes, we only publish a subset of the test dataset in the family and relationships domain (`data/GYAFC/`), the outputs (`outputs/GYAFC/`) of each system (including our model and all baselines) and the corresponding human references (`references/GYAFC/`). If you want to download the train and validation dataset, please follow the guidance at [https://github.com/raosudha89/GYAFC-corpus](https://github.com/raosudha89/GYAFC-corpus). And then, name the corpora of two styles as the yelp dataset. ## Quick Start First of all, you should specify the dataset. For example, for yelp dataset: ``` export DATASET=yelp ``` If you want to use your own datasets, please follow the guidance of next section [Extend to other tasks and datasets](#extend-to-other-tasks-and-datasets). ### Step 1: Pre-train classifier ``` cd classifier python textcnn.py --mode train ``` **Note:** If you get the error `no module named opennmt`, please install `OpenNMT-tf`: `pip install OpenNMT-tf==1.15.0`. ### Step 2: Pre-train two seq2seq (nmt) models using pseudo-parallel data #### 2.1 Prepare pseudo-parallel data To generate pseudo-parallel data, we follow the [template-based method](https://github.com/shentianxiao/language-style-transfer) proposed by [Li et al., 2018](https://aclweb.org/anthology/N18-1169). And we have provided the pseudo-parallel data of the yelp dataset in the `data/yelp/tsf_template` directory. However, if you want to generate the pseudo-parallel data using templates, you can follow [this link](https://github.com/lijuncen/Sentiment-and-Style-Transfer/issues/6) or design your own templates which are suitable for your task and dataset. #### 2.2 Pre-train two seq2seq (nmt) models The default encoder and decoder are **bilstm**. ``` cd nmt python nmt.py --mode train --nmt_direction 0-1 --n_epoch 5 # Pre-train forward (f) model python nmt.py --mode train --nmt_direction 1-0 --n_epoch 5 # Pre-train backward (g) model ``` If you want to adopt **transf
Excerpt of 8,228 characters
Read on GitHubWould you bet a product on this? Bounded 0–100 and slow moving.
matched fp:13d74090aad818ab, llm:Repository topics and description: 'dual-learning, reinceforcement-learning, text-style-transfer, unsupervised-machine-learning' and description: 'A Dual Reinforcement Learning Framework for Unsupervised Text Style Transfer (IJCAI 2019)'.
matched fp:13d74090aad818ab, llm:Repository topics and description: 'dual-learning, reinceforcement-learning, text-style-transfer, unsupervised-machine-learning' and description: 'A Dual Reinforcement Learning Framework for Unsupervised Text Style Transfer (IJCAI 2019)'.
matched fp:13d74090aad818ab, llm:Repository topics and description: 'dual-learning, reinceforcement-learning, text-style-transfer, unsupervised-machine-learning' and description: 'A Dual Reinforcement Learning Framework for Unsupervised Text Style Transfer (IJCAI 2019)'.