NUS Grammatical Error Correction 2017 (SMT-based)

---

This repository contains scripts and models to replicate the results of the state-of-the-art (in 2017) grammatical error correction (GEC) system. The system primarily uses the statistical machine translation (SMT) framework, with neural and character level translation components.

If you use the models from this work, please cite this paper:

@InProceedings{chollampatt2017gec,
  author    = {Chollampatt, Shamil and Ng, Hwee Tou},
  title     = {Connecting the Dots: Towards Human-level Grammatical Error Correction},
  booktitle = {Proceedings of the 12th Workshop on Innovative Use of NLP for Building Educational Applications},
  month     = {September},
  year      = {2017},
  address   = {Copenhagen, Denmark},
  publisher = {Association for Computational Linguistics}
}

Prerequistes

1. cmph - For compact phrase tables
2. nplm - For neural network joint model feature in Moses.
3. Moses Release 3.0
   • Execute ./download.sh in software/ directory to download the specific revision of Moses used in this work.
   • Moses needs to be compiled with ---with-cmph=/path/to/cmph-2.0 --with-nplm=/path/to/nplm --max-kenlm-order=9 this flag

Downloading the Models

Go to models/ and run download.sh download the following model files:

1. Language Models (lm/)

   Model	Filename	Size	Description
   Train LM	concat-train.trie	128 MB	5-gram LM trained using the target side of the
   Wiki LM	wiki.trie	6 GB	5-gram LM trained using English Wikipedia dump (1.78B tokens)
   94B CCLM	94Bcclm.trie	153 GB	5-gram LM trained using subset of CommonCrawl corpus (94B tokens)
   94B CCLM	94Bcclm.filtered.trie	34 GB	94B CCLM filtered according to the phrase-table and the CoNLL-2014 and the development set (faster loading)
   Character LM	charLM.bin	1 MB	Character-level LM for character-level SMT

2. Phrase Tables (pt/)

   Model	Filename	Size	Description
   WordSMT-GEC PT	phrase-table.minphr	851 MB	Phrase table used by word-level SMT
   CharSMT-GEC PT	phrase-table.char.gz	12 MB	Phrase table used by character-level SMT

3. Neural Network Joint Model (NNJM) (nnjm/)

   Model	Filename	Size	Description
   Adapted NNJM	nnjm_adapted.weights	268 MB	Model file with trained weights in NPLM format (compatible with MOSES)
   Vocab	output.vocab, source.vocab, target.vocab	370 KB	Vocabulary files

4. Operation Sequence Model (OSM) (osm/)

   Model	Filename	Size	Description
   Binarized OSM	operationLM.bim	323 MB	Binarized operation sequence model

5. Word Class Language Model (WCLM) (wclm/)

   Model	Filename	Size	Description
   Wiki Word Class LM	wiki.classes.kenlm	13 GB	9-gram word
   Word2Class Mapping	wiki.classes.gen.gz	7.6 MB	Binarized operation sequence model

6. Sparse Weights (sparse/)

   Model	Filename	Size	Description
   Sparse Weights	word.moses.ini.sparse	5.6 MB	Sparse feature weights with 1 word context.

Running the system

To run the system on any test set, execute the following script on the tokenized input file.

./run.sh path/to/input/file /path/to/output/directory

License

The code and models in this repository are licensed under the GNU General Public License Version 3.

For commercial use of this code and models, separate commercial licensing is also available. Please contact:

1. Shamil Chollampatt ([email protected])
2. Hwee Tou Ng ([email protected])