---

THIS IS A DEVELOPMENTAL REPOSITORY FOR M2Scorer
FOR AN OFFICIAL VERSION (VERSION 3.2), visit http://www.comp.nus.edu.sg/~nlp/conll14st.html
OR CHECK OUT THE RELEASES: https://github.com/nusnlp/m2scorer/releases

---

M^2Scorer

This is the scorer for evaluation of grammatical error correction systems. This program is free software: you can redistribute it and/or modify it under the terms of the GNU General Public License (See LICENSE).

If you are using the NUS M^2 scorer in your work, please include a citation of the following paper:

Daniel Dahlmeier and Hwee Tou Ng. 2012. Better Evaluation for Grammatical Error Correction. In Proceedings of the 2012 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (NAACL 2012).

Contents

1. Quickstart
2. Pre-requisites
3. Using the scorer
   3.1 System output format
   3.2 Scorer's gold standard format
4. Converting the CoNLL-2014 data format
5. Revisions
   5.1 Alternative edits
   5.2 F-beta measure
   5.3 Handling of insertion edits
   5.4 Bug fix for scoring against multiple sets of gold edits, and dealing with sequences of insertion/deletion edits

Quickstart

./m2scorer [-v] SYSTEM SOURCE_GOLD 

SYSTEM = the system output in sentence-per-line plain text. SOURCE_GOLD = the source sentences with gold edits.

Pre-requisites

The following dependencies have to be installed to use the M^2 scorer.

• Python (>= 2.6.4, < 3.0, older versions might work but are not tested)
• nltk (http://www.nltk.org, needed for sentence splitting)

Using the scorer

Usage: m2scorer [OPTIONS] SYSTEM SOURCE_GOLD

where
SYSTEM - system output, one sentence per line
SOURCE_GOLD - source sentences with gold token edits

OPTIONS
  -v    --verbose             -  print verbose output
  --very_verbose              -  print lots of verbose output
  --max_unchanged_words N     -  Maximum unchanged words when extracting edits. Default = 2.
  --ignore_whitespace_casing  -  Ignore edits that only affect whitespace and casing. Default no.
  --beta                      -  Set the ratio of recall importance against precision. Default = 0.5.

2.1 System output format

The sentences should be in tokenized plain text, sentence-per-line format.

Format:

<tokenized system output for sentence 1>
<tokenized system output for sentence 2>
 ...

Examples of tokenization:
Original : He said, "We shouldn't go to the place. It'll kill one of us."
Tokenized : He said , " We should n't go to the place . It 'll kill one of us . "

Note: Tokenization in the CoNLL-2014 shared task uses NLTK word tokenizer.

Sample output:
===> system <=== A cat sat on the mat . The Dog .

Scorer's gold standard format

SOURCE_GOLD = source sentences (i.e. input to the error correction system) and the gold annotation in TOKEN offsets (starting from zero).

Format:

S <tokenized system output for sentence 1>
A <token start offset> <token end offset>|||<error type>|||<correction1>||<correction2||..||correctionN|||<required>|||<comment>|||<annotator id>
A <token start offset> <token end offset>|||<error type>|||<correction1>||<correction2||..||correctionN|||<required>|||<comment>|||<annotator id>

S <tokenized system output for sentence 2>
A <token start offset> <token end offset>|||<error type>|||<correction1>||<correction2||..||correctionN|||<required>|||<comment>|||<annotator id>

Notes:

• Each source sentence should appear on a single line starting with "S "
• Each source sentence is followed by zero or more annotations.
• Each annotation is on a separate line starting with "A ".
• Sentences are separated by one or more empty lines.
• The source sentences need to be tokenized in the same way as the system output.
• Start and end offset for annotations are in token offsets (starting from zero).
• The gold edits can include one or more possible correction strings. Multiple corrections should be separate by '||'.
• The error type, required field, and comment are not used for scoring at the moment. You can put dummy values there.
• The annotator ID is used to identify a distinct annotation set by which system edits will be evaluated.
  • Each distinct annotation set, identified by an annotator ID, is an alternative
  • If one sentence has multiple annotator IDs, score will be computed for each annotator.
  • If one of the multiple annotation alternatives is no edit at all, an edit with type 'noop' or with offsets '-1 -1' must be specified.
  • The final score for the sentence will use the set of edits by an annotation set maximizing the score.

Example:

The gold annotation file can be found here: example/source_gold

S The cat sat at mat .
A 3 4|||Prep|||on|||REQUIRED|||-NONE-|||0
A 4 4|||ArtOrDet|||the||a|||REQUIRED|||-NONE-|||0

S The dog .
A 1 2|||NN|||dogs|||REQUIRED|||-NONE-|||0
A -1 -1|||noop|||-NONE-|||-NONE-|||-NONE-|||1

S Giant otters is an apex predator .
A 2 3|||SVA|||are|||REQUIRED|||-NONE-|||0
A 3 4|||ArtOrDet|||-NONE-|||REQUIRED|||-NONE-|||0
A 5 6|||NN|||predators|||REQUIRED|||-NONE-|||0
A 1 2|||NN|||otter|||REQUIRED|||-NONE-|||1

Let the system output, example/system be

A cat sat on the mat .
The dog .
Giant otters are apex predator .

Run the M^2Scorer as follows:

./m2scorer example/system example/source_gold 

The evaluation output will be will be:

Precision   : 0.8000
Recall      : 0.8000
F_0.5       : 0.8000

Explanation: For the first sentence, the system makes two valid edits {(at-> on), (\epsilon -> the)} and one invalid edit (The -> A).

For the second sentence, despite missing one gold edit (dog -> dogs) according to annotation set 0, the system misses nothing according to set 1.

For sentence #3, according to annotation set 0, the system makes two valid edits {(is -> are), (an -> \epsilon)} and misses one edit (predator -> predators); however according to set 1, the system makes two unnecessary edits {(is -> are), (an -> \epsilon)} and misses one edit (otters -> otter).

By the case above, there are four valid edits, one unnecessary edit, and one missing edit. Therefore precision is 4/5 = 0.8. Similarly for recall. In the above example, the beta value for the F-measure is 0.5 (the default value).

###Converting the CoNLL-2014 data format The data format used in the M^2 scorer differs from the format used in the CoNLL-2014 shared task (http://www.comp.nus.edu.sg/~nlp/conll14st.html) in two aspects:

• sentence-level edits
• token edit offsets

To convert source files and gold edits from the CoNLL-2014 format into the M^2 format, run the preprocessing script bundled with the CoNLL-2014 training data.