Skip to main content

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

Project description

Scorch¹

Build Status PyPI

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

With pip

python3 -m pip install scorch

From git

Download from master with

git clone https://github.com/LoicGrobol/scorch.git

Install with

python3 -m pip install .

Alternatively, just running scorch.py without installing should work as long as you have all the dependencies installed

python3 scorch.py -h

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 gold directory are expected to be present in the sys directory, with exactly one sys file for each gold 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-0.0.17.tar.gz (14.9 kB view details)

Uploaded Source

Built Distribution

scorch-0.0.17-py3-none-any.whl (13.5 kB view details)

Uploaded Python 3

File details

Details for the file scorch-0.0.17.tar.gz.

File metadata

  • Download URL: scorch-0.0.17.tar.gz
  • Upload date:
  • Size: 14.9 kB
  • Tags: Source
  • Uploaded using Trusted Publishing? No

File hashes

Hashes for scorch-0.0.17.tar.gz
Algorithm Hash digest
SHA256 0a7c8712fc49fbd57c0b6e0816f789dc1f54b0e6738b069abb982911cd79d442
MD5 6439be889bb56669e3d3d6968dcbca16
BLAKE2b-256 525b88e809d995739d7a5f27bc412194dfab5baf920a81ac7fd23cbf8277ae39

See more details on using hashes here.

File details

Details for the file scorch-0.0.17-py3-none-any.whl.

File metadata

File hashes

Hashes for scorch-0.0.17-py3-none-any.whl
Algorithm Hash digest
SHA256 839c10bd92761ce638e94517162c709c7783790c16784e1985a23d334684561c
MD5 141731d33642ef5eb26fc3afbdd968ad
BLAKE2b-256 b6fb0b52faf381905f3b6ac715bcce9e44c199a0ec7de631e5f820a574bd2d30

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