This repository contains baseline models, training scripts, and instructions on how to reproduce our results for our state-of-art grammar correction system from M. Junczys-Dowmunt, R. Grundkiewicz: Phrase-based Machine Translation is State-of-the-Art for Automatic Grammatical Error Correction, EMNLP 2016.
@InProceedings{junczysdowmunt-grundkiewicz:2016:EMNLP2016,
author = {Junczys-Dowmunt, Marcin and Grundkiewicz, Roman},
title = {Phrase-based Machine Translation is State-of-the-Art for
Automatic Grammatical Error Correction},
booktitle = {Proceedings of the 2016 Conference on Empirical Methods in
Natural Language Processing},
month = {November},
year = {2016},
address = {Austin, Texas},
publisher = {Association for Computational Linguistics},
pages = {1546--1556},
url = {https://aclweb.org/anthology/D16-1161}
}
Last update: 3/8/2018
The train-2018
directory contains updated training scripts and instructions
that we used to create SMT systems in our paper: R. Grundkiewicz, M.
Junczys-Dowmunt Near Human-Level Performance in Grammatical Error Correction
with Hybrid Machine Translation, NAACL
2018 [bibtex]
Main modifications include switching to NLTK tokenization, using BPE subword
segmentation, and adding GLEU tuning.
We publish the text data we used for training a web-scale language model (CCLM): http://data.statmt.org/romang/gec-emnlp16/sim/
The data is tokenized with NLTK tokenizer and truecased with the Moses
truecaser. All parts are separate files and can be separately extracted using
the xz
tool. Parts 00-04 consist of 1000M lines, part 05 consists of
291,262,763 lines.
Outputs generated by our models for the JFLEG data sets are available in the
folder jfleg
. These are produced by our systems tuned on M^2. See the README
in that folder for more information.
New: We also report results for the systems tuned on GLEU using JFLEG dev.
The phrase table that we have made publicly available for download were filtered for CoNLL test sets. The evaluation of our systems with that PT on other data sets makes no sense. Now, we provide the original unfiltered phrase table in binarized format (due to its size). The outputs for CoNLL test sets produced with a binarized PT should remain unchanged.
All .ini files and instructions how to use them have been updated.
We have updated the model which use CCLM and sparse features. That model was
used to generate results reported in the paper as Best sparse + CCLM. Moses
.ini files are available in the folder models
.
We also provide the script models/run_gecsmt.py
to run our models (see notes
below).
The results reported in the camera-ready version of the paper on the dataset from Bryant and Ng (2015) (Tab. 4, three last columns) are understated due to the invalid preparation of the M2 file. The correct scores are as follows:
System | Prec. | Recall | M^2 |
---|---|---|---|
Baseline | 69.22 | 37.00 | 58.95 |
+CCLM | 76.66 | 36.39 | 62.77 |
Best dense | 71.11 | 37.44 | 60.27 |
+CCLM | 79.76 | 39.52 | 66.27 |
Best sparse | 76.48 | 35.99 | 62.43 |
+CCLM | 80.57 | 39.74 | 66.83 |
We would like to thank Shamil Chollampatt for reporting this issue!
Outputs generated by our models for the CoNLL-2014 test set are available in
the folder outputs
. These correspond to Table 4 of our paper. See the README
in that folder for more information.
You can download and run our baseline models (1,3G).
models/
├── data
│ ├── lm.cor.kenlm
│ ├── osm.kenlm
│ ├── phrase-table.0-0.gz
│ └── phrase-table.0-0.unfiltered.minphr
├── moses.dense-cclm.mert.avg.ini
├── moses.dense.mert.avg.ini
├── moses.sparse-cclm.mert.avg.ini
├── moses.sparse.mert.avg.ini
└── sparse
├── moses.cc.sparse
└── moses.wiki.sparse
The four configuration *.ini
files corresponds to the last four systems
described in Table 4.
To use the models you need to install Moses
decoder (branch master
). It has
to be compiled with support for 9-gram kenLM language models, and binarized
tables by providing path to CMPH library (see details
here), e.g.:
/usr/bin/bjam -j16 --max-kenlm-order=9 --with-cmph=/path/to/cmph
The language model data are available in separate packages:
- Wikipedia language model (22G)
- Common Crawl language model (26G)
The packages contain:
wikilm/
├── wiki.blm
├── wiki.classes.gz
└── wiki.wclm.kenlm
cclm/
├── cc.classes.gz
├── cc.kenlm
└── cc.wclm.kenlm
Adjust absolute paths in moses.*.ini
files. You can do this by replacing
/path/to/
with the path to the directory where you downloaded models and
language models. Finally, run moses, e.g.:
/path/to/mosesdecoder/bin/moses -f moses.dense.mert.avg.ini < input.txt
The input file should contain one sentence per line and each sentence has to
follow the Moses tokenization and truecasing as it is presented in
train/run_cross.perl
.
Alternatively you can use the script models/run_gecsmt.py
, which performs
pre- and postprocessing, e.g.:
python ./run/run_gecsmt.py -f moses.ini -w workdir -i input.txt -o output.txt
It can be used to evaluate M2 input:
python ./run/run_gecsmt.py -f moses.ini -w workdir -i test2014.m2 --m2
You will need to provide paths to Moses, Lazy and this repository. Use --help
option for more details.
Running our models might give slightly different results (up to +/- 0.0020 F-score) than the results presented in the paper due to the different versions of the official CoNLL-2014 test set (we used the version provided during the CoNLL shared task), M2Scorer, NLTK tokenizer, Moses, and the LM used for truecasing.
Training is described in the README in the folder train
.
This project was partially funded by the Polish National Science Centre (Grant No. 2014/15/N/ST6/02330).