Low Resource Grammatical Error Correction Using Wikipedia Edits

This repository contains the code and data for: Low Resource Grammatical Error Correction Using Wikipedia Edits

Please use the following citation:

@InProceedings{boyd2018wnut,
  author    = {Adriane Boyd},
  title     = {Using Wikipedia Edits in Low Resource Grammatical Error Correction},
  booktitle = {Proceedings of the 4th Workshop on Noisy User-generated Text},
  publisher = {Association for Computational Linguistics},
  year      = {2018},
  url       = {http://aclweb.org/anthology/W18-6111}
}

To replicate the experiments in the paper, install mlconvgec2018 and download the data and models as described in the following section.

To generate grammatical error correction data using the method presented in the paper, see Generating GEC Data with Wiki Edits and ERRANT.

Install mlconvgec2018

Clone mlconvgec2018:

git clone https://github.com/nusnlp/mlconvgec2018/

The experiments were performed using commit 95725921.

Follow installation instructions in mlconvgec2018/README.md.

Install moses (the experiments used commit 03578921c) and m2scorer 2.3.

Download Data

Download the Falko-MERLIN GEC Corpus and German Wikipedia data: data.tar.gz

Run ./prepare-training-data.sh in data/ to generate the training files for the experiment.

Falko / MERLIN Corpora

The original corpora are available at:

• Falko
• MERLIN

Tables linking the Falko/MERLIN sentence pairs to their text IDs from the original corpora are in data/source/. For both corpora, the ctok (original) and ZH1/TH1 (correction) layers were extracted for the Falko-MERLIN GEC Corpus.

Data Licenses

• Falko: CC BY 3.0
• MERLIN: CC BY-SA 4.0
• Wikipedia: CC BY-SA 3.0

Download Models

Download the BPE and embeddings models: models.tar.gz

Due to its size, the language model is not provided for download.

Run Experiment

Copy scripts/run_exp.sh to mlconvgec2018/training/ and adjust the paths and settings at the beginning of the script.

Adjust the paths in mlconvgec2018/paths.sh.

The script run_exp.sh is used with a single argument to run the preprocessing, training, and evaluation steps for one training data set, e.g.:

./run_exp.sh fm-train-wiki.100K

where fm-train-wiki.100K.src and fm-train-wiki.100K.trg can be found in DATA_DIR in paths.sh.

The corresponding test output is in a timestamped directory, e.g.:

fm-train-wiki.100K-2018-08-27-0935-test-dim500-bpe30000-seed1000

under:

output.tok.txt
output.reranked.eo.tok.txt
output.reranked.eolm-lm.binary.tok.txt
output.reranked.lm-lm.binary.tok.txt
output.reranked.lmnorm-lm.binary.tok.txt

and the m2scorer evaluations in the corresponding .m2scorer files:

output.tok.txt.m2scorer
output.reranked.eo.tok.txt.m2scorer
output.reranked.eolm-lm.binary.tok.txt.m2scorer
output.reranked.lm-lm.binary.tok.txt.m2scorer
output.reranked.lmnorm-lm.binary.tok.txt.m2scorer

Generating GEC Data with Wiki Edits and ERRANT

The following sections describe how German Wikipedia edits were extracted from the Wikipedia revision history and filtered using ERRANT.

If necessary, initialize the submodules containing ERRANT and WikiEdits:

git submodule update --init --recursive

Wiki Edits

Sentence pairs with small numbers of edits were extracted from Wikipedia revision history using Wiki Edits with minor modifications to support German available in the submodule wikiedits.

Extract Wikipedia edits from the Wikipedia history dumps with a maximum of 60 words per sentence, e.g.:

7zr e -so dewiki-20180601-pages-meta-history1.xml-p1p3342.7z | ./bin/wiki_edits.py -l german -t --max-words 60 > dewiki-20180601-pages-meta-history1.xml-p1p3342.wikiedits.60

ERRANT

Install ERRANT

The submodule errant contains the extensions to ERRANT for German that were used for these experiments.

Install spacy and download the German models:

pip install -U spacy==2.0.11
python -m spacy download de

Install TreeTagger under errant/resources/tree-tagger-3.2 and copy the German model german-utf8.par (download: german-par-linux-3.2-utf8.bin.gz) to errant/resources/tree-tagger-3.2/lib/.

Install treetaggerwrapper:

pip install treetaggerwrapper

Using ERRANT for German

Profile the tokenized fm-train data, specifying German with the new option -lang de:

python parallel_to_m2.py -orig fm-train.src -cor fm-train.trg -out fm-train.m2 -lang de

Analyze untokenized Wikipedia edits using the same script with the new option -tok to have ERRANT also tokenize the input data with the spacy tokenizer:

python parallel_to_m2.py -orig wiki-unfiltered.src -cor wiki-unfiltered.trg -out wiki-unfiltered.m2 -lang de -tok

Filter the Wikipedia edits with reference to the Falko/MERLIN training data:

python filter_m2.py -filt wiki-unfiltered.m2 -ref fm-train.m2 -out wiki-filtered.m2

Convert the filtered wiki m2 back to a plaintext file of parallel sentences:

python m2_to_parallel.py -m2 wiki-filtered.m2 -out wiki-filtered.src-trg.txt