mmax2conll

Script to convert coreference data in MMAX format to CoNLL format or raw text files.

See CoNLL-specification.md and MMAX-specification.md for extensive descriptions of the CoNLL and MMAX formats.

Usage

mmax2conll.py

Because the COREA corpus saves its sentence information in the *_words.xml files but the SoNaR-1 corpus saves this separately in *_sentence_level.xml files, specifying a sentences file is optional.

To automatically find all (sub)folders that contain a Basedata and Markables folder as direct children and convert all data in those folders, run:

python -m mmax2conll path/to/config.yml path/to/output_dir -d path/to/some/folder [-d path/to/another/folder ...]

To only convert one pair (or triple) of files, run:

python -m mmax2conll path/to/config.yml path/to/output.conll path/to/some_words.xml path/to/a_coref_level.xml [path/to/a_sentence_level.xml]

mmax2raw.py

To automatically find all (sub)folders that contain a Basedata and Markables folder as direct children and convert all data in those folders, run:

mmax2raw.py path/to/output_dir -d path/to/some/folder [-d path/to/another/folder ...]

To only convert one file, run:

mmax2raw.py path/to/output.txt path/to/some_words.xml

Columns of CoNLL output

These scripts were first used to convert data from the COREA (Ch.7 p.115 -- 128) dataset to CoNLL and COREA does not contain the following information:

• POS tags
• constituency tree
• predicates
• speaker/author information
• named entities

Therefore these scripts strictly do not output data in CoNLL format. The following values and place-holders are used.

[1]: The part numbers of DCOI start at 1, where the part numbers in a CoNLL file start at 0. To keep the origin of the data clear this 1-based part number is not changed, but instead an empty part 0 is added to those files.

[2]: The reference spans are not closed in the correct order if they end at the same word. The following is an example of output from mmax2conll:

          (10
            -
      (52|(55
          52)
            -
10)|55)|(133)

While pedantically correct would be:

          (10
            -
      (55|(52
          52)
            -
(133)|55)|10)

Issues

• Skipping a whole file one any error is too wasteful
• 'on_missing' config key is not validated before use
• basedata_dir and markables_dir should not be configuration keys
• Too many methods in main.py are marked as class methods

References

Christoph Müller and Michael Strube.
Multi-Level Annotation in MMAX
In Proceedings of the 4th SIGDIAL Workshop. 2003.
URL http://www.speech.cs.cmu.edu/sigdial2003/proceedings/07_LONG_strube_paper.pdf

Iris Hendrickx, Gosse Bouma, Walter Daelemans and Véronique Hoste.
COREA: Coreference Resolution for Extracting Answers for Dutch
Essential Speech and Language Technology for Dutch, Ch.7, p.115 -- 128. 2013.
Editors: Peter Spyns, Jan Odijk
https://link.springer.com/book/10.1007/978-3-642-30910-6
10.1007/978-3-642-30910-6

SoNaR: https://ivdnt.org/downloads/taalmaterialen/tstc-sonar-corpus