Skip to main content

Alternative scorer for the CoNLL-2011/2012 shared tasks on coreference resolution.

Project description

Scorch¹

Build Status PyPI Total alerts

This is an alternative implementation of the coreference scorer for the CoNLL-2011/2012 shared tasks on coreference resolution.

It aims to be more straightforward than the reference implementation, while maintaining as much compatibility with it as possible.

The implementations of the various scores are as close as possible from the formulas used by Pradhan et al. (2014), with the edge cases for BLANC taken from Recasens and Hovy (2011).


1. Scorer for coreference chains.

Usage

scorch gold.json sys.json out.txt

Install

From the cheeseshop

python3 -m pip install --user scorch

Or directly from git

python3 -m pip install --user git+https://github.com/LoicGrobol/scorch.git

Formats

Single document

The input files should be JSON files with a "type" key at top-level

  • If "type" is "graph", then top-level should have at top-level
    • A "mentions" key containing a list of all mention identifiers
    • A "links" key containing a list of pairs of corefering mention identifiers
  • If "type" is "clusters", then top-level should have a "clusters" key containing a mapping from clusters ids to cluster contents (as lists of mention identifiers).

Of course the system and gold files should use the same set of mention identifiers…

Multiple documents

If the inputs are directories, files with the same base name (excluding extension) as those present in the sys directory are expected to be present in the gold directory, with exactly one gold file for each sys file. In that case, the output scores will be the micro-average of the individual files scores, ie their arithmetic means weighted by the relative numbers of

  • Gold mentions for Recall
  • System mentions for Precision
  • The sum of the previous two for F₁

This is different from the reference interpretation where

  • MUC weighting ignores mentions in singleton entities
    • This should not make any difference for the CoNLL-2012 dataset, since singleton entities are not annotated.
    • For datasets with singletons, the shortcomings of MUC are well known, so this score shouldn't matter much
  • BLANC is calculated by micro-averaging coreference and non-coreference separately, using the number of links as weights instead of the number of mentions.
    • This is roughly equivalent to weighting coreference scores per document by their number of non-singleton clusters and non-coreference scores by the square of their number of mentions. This give disproportionate importance to large documents, which is not desirable in heterogenous corpora

The CoNLL average score is the arithmetic mean of the global MUC, B³ and CEAFₑ F₁ scores.

Sources

License

Unless otherwise specified (see below), the following licence (the so-called “MIT License”) applies to all the files in this repository. See also LICENSE.md.

Copyright 2018 Loïc Grobol <loic.grobol@gmail.com>

Permission is hereby granted, free of charge, to any person obtaining a copy of this software and
associated documentation files (the "Software"), to deal in the Software without restriction,
including without limitation the rights to use, copy, modify, merge, publish, distribute,
sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all copies or
substantial portions of the Software.

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT
NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM,
DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT
OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.

License exceptions

Project details


Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

scorch-0.0.22.tar.gz (14.7 kB view details)

Uploaded Source

Built Distribution

scorch-0.0.22-py3-none-any.whl (14.6 kB view details)

Uploaded Python 3

File details

Details for the file scorch-0.0.22.tar.gz.

File metadata

  • Download URL: scorch-0.0.22.tar.gz
  • Upload date:
  • Size: 14.7 kB
  • Tags: Source
  • Uploaded using Trusted Publishing? No
  • Uploaded via: twine/1.13.0 pkginfo/1.5.0.1 requests/2.21.0 setuptools/40.8.0 requests-toolbelt/0.9.1 tqdm/4.31.1 CPython/3.7.1

File hashes

Hashes for scorch-0.0.22.tar.gz
Algorithm Hash digest
SHA256 3b6b39af4d9d404c352dc3aaea1e58d90ec436343afc21abb07f52b5cd71f5bb
MD5 42e2b2267db1e4de0581f9854b9d3751
BLAKE2b-256 04ba512aa7cc04f0232e2f08e8357f8268c594cac36430cbc15a9abe435e3870

See more details on using hashes here.

File details

Details for the file scorch-0.0.22-py3-none-any.whl.

File metadata

  • Download URL: scorch-0.0.22-py3-none-any.whl
  • Upload date:
  • Size: 14.6 kB
  • Tags: Python 3
  • Uploaded using Trusted Publishing? No
  • Uploaded via: twine/1.13.0 pkginfo/1.5.0.1 requests/2.21.0 setuptools/40.8.0 requests-toolbelt/0.9.1 tqdm/4.31.1 CPython/3.7.1

File hashes

Hashes for scorch-0.0.22-py3-none-any.whl
Algorithm Hash digest
SHA256 b308a8c71c432f4ef6fbb66998ab5d33881990649fc487bdd89b941cc89dbad8
MD5 64c1b5fbf25a157a6980131cb5e7a0fb
BLAKE2b-256 dd71c12715d3886afcf171ddb07fbe9043c852ac9b99be7645fa50625fdc3766

See more details on using hashes here.

Supported by

AWS Cloud computing and Security Sponsor Datadog Monitoring Fastly CDN Google Download Analytics Pingdom Monitoring Sentry Error logging StatusPage Status page