Spelling Corrector Experimental models

This is experimental project for research of spelling corrector models. We've researched several configurations of systems for spelling corrections:

Language model (all based on ELMO model):

• TFKuznetsovELMO (Best by quality but slow)
• PyTorchELMO (Good for speed and good for quality)
• TransformerELMO (Nice for speed but bad quality)
• RusVecores ELMO (Nice for speed but bad quality)
• TFHubELMO (Not good solution)

Error model and candidates generator:

• Levenshtein
• Weighted Levenshtein

Reranker for corrections making:

• Summator (decision selected for hypotheses with maximal sum of language_model score and error_score)
• Logistic Regression (decision is made by trainable Regression algorithm which uses multiple features includin language model score and error score and many other)

##Root documentation

http://192.168.10.188:8081/index.php/Spelling_Corrector

Table with comparison of models on different datasets

https://docs.google.com/spreadsheets/d/1QzYF2O5z1nQR8gic8uFTx4TnB242EV3uNXPjl-xTppg/edit#gid=0

Basic usage

How to prepare spelling corrector object with specific implementation underneath

git clone https://github.com/acriptis/spelling_corrector_experiments

pip install -r requirements.txt

then you can launch tests:

python tests/test_spelling_correctors.py

but it requires about 30GB of RAM and loads disck space (model ELMO40inKuz consu,es about 4.5GB of disk space, TorchELMO40in model consumes about 5GB of disk space) (But it loads all models, )

As-Server usage

TODO How to run spelling corrector as server

• deepepavlov riseapi?
• django rest api service
• flask
• TODO wrap solution to Docker image?

How to train your own components

TODO

Language model

TODO

ReRanker

TODO

Spelling Corrector Candidates Generator

TODO describe how to configure