Implementation of Stanford multi-seive coreference resolution approach for Dutch. This is a draft version of the code. An official first release will be made available on github/cltl upon completion and basic testing of the first version of the system.
The current implementation works on naf input files parsed by Alpino (i.e. it works for Dutch).
Future plans:
- separate Alpino specific functions from general naf-extraction functions (extend to other languages)
- create library for English
From command line:
$ python multisieve_coreference < inputfile.naf
From python:
from multisieve_coreference import process_coreference
process_coreference(naf_object)
Calling process_coreference will change the naf_object in-place by adding coref nodes (if any).
Gaps in mention spans (mostly left-out punctuation marks) are not filled by default. To make sure mentions only refer to consecutive spans, pass -f or --fill-gaps on the command line or call process_coreference(naf_object, fill_gaps=True).
!! NB !! Singleton clusters are left out by default. To Include singleton clusters pass -s or --include_singletons on the command line or call process_coreference(naf_object, include_singletons=True).
- Mentions are not ordered at all, in contrast to the description of the algorithm by Lee et al. (2013)
- Only one mention of the current coreference classes should be considered as antecedent candidate. See 3.2.1 Mention Selection in a Given Sieve in Lee et al. (2013)
- Mention attributes are not shared among mentions in the same coreference class, in contrast to the description of the algorithm by Lee et al. (2013)
- Alpino uses two types of dependencies ("deep" and "shallow"). Make sure these are handled correctly.
-
global stop_wordsshould be aset(and not a global) and doesn't seem to be used consistently. -
linguisticProcessorslayer should be added tonafHeader -
create_mentiondocstring -
femandmascdo not appear in output of Alpino, but are used to identify gender - Many variables and functions use
idin their name while they actually contain or use offsets. - Duplicate code in
resolve_relative_pronoun_structuresandresolve_reflective_pronoun_structures -
get_predicative_informationseems to miss some constructs -
dep2headsshouldn't need lists as values because a dependency tree is a tree. - Add punctuation marks to mention spans in post processing if they are in the middle of the mention. They aren't in the span in the first place because punctuation is filtered out of the dependency tree to make sure punctuation that isn't in the middle of a mention is not included.
-
post_processshould only allow adding punctuation in the gaps of a span -
Cmention.coreference_prohibiteddoes not seem to be used. -
check_if_quotation_contains_dependentis a mess.
Instead of passing around a bare dictionary of mentions, an iterable MentionCollection object would be a better idea because it would be able to take care of mention ordering and filtering.
It would be great if all changing information (mostly related to which mentions should or should not be in the same coreference class) is kept in one object, instead of spread out amongst the mention objects themselves. (This is currently the case with Cmention.coreference_prohibited.)
It would be even nicer if every sieve would inherit from some Sieve abstraction and these sieves could be easily plugged and played by moving them around in some input sequence.
- Antske Fokkens
- [email protected]
- [email protected]
- http://antskefokkens.info
- Vrije University of Amsterdam
- Martin van Harmelen
- [email protected]
- [email protected]
- Vrije Universiteit Amsterdam
Sofware distributed under Apache License v2.0, see LICENSE file for details.
Heeyoung Lee, Angel Chang, Yves Peirsman, Nathanael Chambers, Mihai Surdeanu, and Dan Jurafsky. 2013. Deterministic coreference resolution based on entity-centric, precision-ranked rules. Comput. Linguist. 39, 4 (December 2013), 885-916. DOI=http://dx.doi.org/10.1162/COLI_a_00152