This is the code repo for the UIST 2019 paper Type, Then Correct: Intelligent Text Correction Techniques for Mobile Text Entry Using Neural Networks. It includes the network structure, the training/testing/deploy files, and the data-processing files.

To get the training data for the network, we used the CoNLL correction tasks data, year 13-14. You can go to DataProcess/CoNLL to check out the related processing code.

We also used the Yelp+Amazon review data, you can find it here (related github project: https://github.com/nhviet1009/Character-level-cnn-pytorch). For those datasets, because they're usually good text without errors, we performed perturbation (injecting errors). The details could be found under DataProcess/PerturbNormalDataset

The training format of the data is provided in the example file under DataProcess. Each training example is composed of text with errors plus the correction, and the expected output. For the output format, please refer to our paper for more details.

python Train.py --train small_train_data5_amazon --batch_size 128 --test small_train_data5_amazon --elr 1e-4 --dlr 5e-4 --epochs 20 -teacher 0.5 --test_freq 1 --dropout 0.2 --clip 10 --hidden_size 300 --load_en best_en_5out_amazon.pth  --only_lowercase 1

small_train_data5_amazon is our training data file; elr/dlr is encoder/decoder learning rate; teacher is the teacher rate.

You can find our trained model and the training data here

We also provided a script for you to deploy this correction algorithm on servers. You can use HTTP protocol to make requests & responses.

Here's a video demo

Please install numpy, nltk, symspellpy, Beautiful Soup 4 for data processing

And the Neural Network file uses Pytorch version 0.4.1.

If you use the code in your paper, then please cite it as:

@inproceedings{Zhang:2019:TCI:3332165.3347924,
 author = {Zhang, Mingrui Ray and Wen, He and Wobbrock, Jacob O.},
 title = {Type, Then Correct\&\#58; Intelligent Text Correction Techniques for Mobile Text Entry Using Neural Networks},
 booktitle = {Proceedings of the 32Nd Annual ACM Symposium on User Interface Software and Technology},
 series = {UIST '19},
 year = {2019},
 isbn = {978-1-4503-6816-2},
 location = {New Orleans, LA, USA},
 pages = {843--855},
 numpages = {13},
 url = {http://doi.acm.org/10.1145/3332165.3347924},
 doi = {10.1145/3332165.3347924},
 acmid = {3347924},
 publisher = {ACM},
 address = {New York, NY, USA},
 keywords = {gestures, natural language processing, text editing, touch},
}