This repository contains the grammatical ERRor ANnotation Toolkit (ERRANT) described in:
Christopher Bryant, Mariano Felice, and Ted Briscoe. 2017. Automatic annotation and evaluation of error types for grammatical error correction. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). Vancouver, Canada.
Mariano Felice, Christopher Bryant, and Ted Briscoe. 2016. Automatic extraction of learner errors in ESL sentences using linguistically enhanced alignments. In Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers. Osaka, Japan.
If you make use of this code, please cite the above papers. More information about ERRANT can be found here. In particular, see Chapter 5 for definitions of error types.
The main aim of ERRANT is to automatically annotate parallel English sentences with error type information. Specifically, given an original and corrected sentence pair, ERRANT will extract the edits that transform the former to the latter and classify them according to a rule-based error type framework. This can be used to standardise parallel datasets or facilitate detailed error type evaluation. Annotated output files are in M2 format and an evaluation script is provided.
Original: This are gramamtical sentence .
Corrected: This is a grammatical sentence .
Output M2:
S This are gramamtical sentence .
A 1 2|||R:VERB:SVA|||is|||REQUIRED|||-NONE-|||0
A 2 2|||M:DET|||a|||REQUIRED|||-NONE-|||0
A 2 3|||R:SPELL|||grammatical|||REQUIRED|||-NONE-|||0
A -1 -1|||noop|||-NONE-|||REQUIRED|||-NONE-|||1
In M2 format, a line preceded by S denotes an original sentence while a line preceded by A indicates an edit annotation. Each edit line consists of the start and end token offset of the edit, the error type, and the tokenized correction string. The next two fields are included for historical reasons (see the CoNLL-2014 shared task) while the last field is the annotator id.
A "noop" edit is a special kind of edit that explicitly indicates an annotator/system made no changes to the original sentence. If there is only one annotator, noop edits are optional, otherwise a noop edit should be included whenever at least 1 out of n annotators considered the original sentence to be correct. This is something to be aware of when combining individual M2 files, as missing noops can affect evaluation.
The easiest way to install ERRANT and its dependencies is using pip. We also recommend installing it in a clean virtual environment (e.g. with venv). The latest version of ERRANT only supports Python >= 3.6.
python3 -m venv errant_env
source errant_env/bin/activate
pip3 install -U pip setuptools wheel
pip3 install errant
python3 -m spacy download en
This will create and activate a new python3 environment called errant_env in the current directory. pip will then update some setup tools and install ERRANT, spaCy, rapidfuzz and spaCy's default English model in this environment. You can deactivate the environment at any time by running deactivate, but must remember to activate it again whenever you want to use ERRANT.
ERRANT was originally designed to work with spaCy v1.9.0 and works best with this version. SpaCy v1.9.0 does not work with Python >= 3.7 however, and so we were forced to update ERRANT to be compatible with spaCy 2. Since spaCy 2 uses a neural system to trade speed for accuracy, this means ERRANT v2.2 is ~4x slower than ERRANT v2.1. We have not yet extended ERRANT to work with spaCy 3, but preliminary tests suggest ERRANT will become even slower.
Consequently, we recommend ERRANT v2.1.0 if speed is a priority and you can use Python < 3.7.
pip3 install errant==2.1.0
ERRANT v2.0.0 was designed to be fully compatible with the BEA-2019 Shared Task. If you want to directly compare against the results in the shared task, you should make sure to install ERRANT v2.0.0 as newer versions may produce slightly different scores. You can also use Codalab to evaluate anonymously on the shared task datasets. ERRANT v2.0.0 is not compatible with Python >= 3.7.
pip3 install errant==2.0.0
If you prefer to install ERRANT from source, you can instead run the following commands:
git clone https://github.com/chrisjbryant/errant.git
cd errant
python3 -m venv errant_env
source errant_env/bin/activate
pip3 install -U pip setuptools wheel
pip3 install -e .
python3 -m spacy download en
This will clone the github ERRANT source into the current directory, build and activate a python environment inside it, and then install ERRANT and all its dependencies. If you wish to modify ERRANT code, this is the recommended way to install it.
Three main commands are provided with ERRANT: errant_parallel, errant_m2 and errant_compare. You can run them from anywhere on the command line without having to invoke a specific python script.
-
errant_parallelThis is the main annotation command that takes an original text file and at least one parallel corrected text file as input, and outputs an annotated M2 file. By default, it is assumed that the original and corrected text files are word tokenised with one sentence per line.
Example:errant_parallel -orig <orig_file> -cor <cor_file1> [<cor_file2> ...] -out <out_m2> -
errant_m2This is a variant of
errant_parallelthat operates on an M2 file instead of parallel text files. This makes it easier to reprocess existing M2 files. You must also specify whether you want to use gold or auto edits; i.e.-goldwill only classify the existing edits, while-autowill extract and classify automatic edits. In both settings, uncorrected edits and noops are preserved.
Example:errant_m2 {-auto|-gold} m2_file -out <out_m2> -
errant_compareThis is the evaluation command that compares a hypothesis M2 file against a reference M2 file. The default behaviour evaluates the hypothesis overall in terms of span-based correction. The
-cat {1,2,3}flag can be used to evaluate error types at increasing levels of granularity, while the-dsor-dtflag can be used to evaluate in terms of span-based or token-based detection (i.e. ignoring the correction). All scores are presented in terms of Precision, Recall and F-score (default: F0.5), and counts for True Positives (TP), False Positives (FP) and False Negatives (FN) are also shown.
Examples:errant_compare -hyp <hyp_m2> -ref <ref_m2> errant_compare -hyp <hyp_m2> -ref <ref_m2> -cat {1,2,3} errant_compare -hyp <hyp_m2> -ref <ref_m2> -ds errant_compare -hyp <hyp_m2> -ref <ref_m2> -ds -cat {1,2,3}
All these scripts also have additional advanced command line options which can be displayed using the -h flag.
As of v2.0.0, ERRANT now also comes with an API.
import errant
annotator = errant.load('en')
orig = annotator.parse('This are gramamtical sentence .')
cor = annotator.parse('This is a grammatical sentence .')
edits = annotator.annotate(orig, cor)
for e in edits:
print(e.o_start, e.o_end, e.o_str, e.c_start, e.c_end, e.c_str, e.type)
errant.load(lang, nlp=None)
Create an ERRANT Annotator object. The lang parameter currently only accepts 'en' for English, but we hope to extend it for other languages in the future. The optional nlp parameter can be used if you have already preloaded spacy and do not want ERRANT to load it again.
import errant
import spacy
nlp = spacy.load('en')
annotator = errant.load('en', nlp)
An Annotator object is the main interface for ERRANT.
annotator.parse(string, tokenise=False)
Lemmatise, POS tag, and parse a text string with spacy. Set tokenise to True to also word tokenise with spacy. Returns a spacy Doc object.
annotator.align(orig, cor, lev=False)
Align spacy-parsed original and corrected text. The default uses a linguistically-enhanced Damerau-Levenshtein alignment, but the lev flag can be used for a standard Levenshtein alignment. Returns an Alignment object.
annotator.merge(alignment, merging='rules')
Extract edits from the optimum alignment in an Alignment object. Four different merging strategies are available:
- rules: Use a rule-based merging strategy (default)
- all-split: Merge nothing: MSSDI -> M, S, S, D, I
- all-merge: Merge adjacent non-matches: MSSDI -> M, SSDI
- all-equal: Merge adjacent same-type non-matches: MSSDI -> M, SS, D, I
Returns a list of Edit objects.
annotator.classify(edit)
Classify an edit. Sets the edit.type attribute in an Edit object and returns the same Edit object.
annotator.annotate(orig, cor, lev=False, merging='rules')
Run the full annotation pipeline to align two sequences and extract and classify the edits. Equivalent to running annotator.align, annotator.merge and annotator.classify in sequence. Returns a list of Edit objects.
import errant
annotator = errant.load('en')
orig = annotator.parse('This are gramamtical sentence .')
cor = annotator.parse('This is a grammatical sentence .')
alignment = annotator.align(orig, cor)
edits = annotator.merge(alignment)
for e in edits:
e = annotator.classify(e)
annotator.import_edit(orig, cor, edit, min=True, old_cat=False)
Load an Edit object from a list. orig and cor must be spacy-parsed Doc objects and the edit must be of the form: [o_start, o_end, c_start, c_end(, type)]. The values must be integers that correspond to the token start and end offsets in the original and corrected Doc objects. The type value is an optional string that denotes the error type of the edit (if known). Set min to True to minimise the edit (e.g. [a b -> a c] = [b -> c]) and old_cat to True to preserve the old error type category (i.e. turn off the classifier).
import errant
annotator = errant.load('en')
orig = annotator.parse('This are gramamtical sentence .')
cor = annotator.parse('This is a grammatical sentence .')
edit = [1, 2, 1, 2, 'SVA'] # are -> is
edit = annotator.import_edit(orig, cor, edit)
print(edit.to_m2())
An Alignment object is created from two spacy-parsed text sequences.
alignment.orig
alignment.cor
The spacy-parsed original and corrected text sequences.
alignment.cost_matrix
alignment.op_matrix
The cost matrix and operation matrix produced by the alignment.
alignment.align_seq
The first cheapest alignment between the two sequences.
An Edit object represents a transformation between two text sequences.
edit.o_start
edit.o_end
edit.o_toks
edit.o_str
The start and end offsets, the spacy tokens, and the string for the edit in the original text.
edit.c_start
edit.c_end
edit.c_toks
edit.c_str
The start and end offsets, the spacy tokens, and the string for the edit in the corrected text.
edit.type
The error type string.
edit.to_m2(id=0)
Format the edit for an output M2 file. id is the annotator id.
If you want to develop ERRANT for other languages, you should mimic the errant/en directory structure. For example, ERRANT for French should import a merger from errant.fr.merger and a classifier from errant.fr.classifier that respectively have equivalent get_rule_edits and classify methods. You will also need to add 'fr' to the list of supported languages in errant/__init__.py.
If you have any questions, suggestions or bug reports, you can contact the authors at:
christopher d0t bryant at cl.cam.ac.uk
mariano d0t felice at cl.cam.ac.uk