license: mit
tags:
- coreference-resolution
language:
- ind
indocoref
Dataset contains articles from Wikipedia Bahasa Indonesia which fulfill these conditions:
The pages contain many noun phrases, which the authors subjectively pick: (i) fictional plots, e.g., subtitles for films,
TV show episodes, and novel stories; (ii) biographies (incl. fictional characters); and (iii) historical events or important events.
The pages contain significant variation of pronoun and named-entity. We count the number of first, second, third person pronouns,
and clitic pronouns in the document by applying string matching.We examine the number
of named-entity using the Stanford CoreNLP
NER Tagger (Manning et al., 2014) with a
model trained from the Indonesian corpus
taken from Alfina et al. (2016).
The Wikipedia texts have length of 500 to
2000 words.
We sample 201 of pages from subset of filtered
Wikipedia pages. We hire five annotators who are
undergraduate student in Linguistics department.
They are native in Indonesian. Annotation is carried out using the Script d’Annotation des Chanes
de Rfrence (SACR), a web-based Coreference resolution annotation tool developed by Oberle (2018).
From the 201 texts, there are 16,460 mentions
tagged by the annotators
Dataset Usage
Run pip install nusacrowd before loading the dataset through HuggingFace's load_dataset.
Citation
@inproceedings{artari-etal-2021-multi,
title = {A Multi-Pass Sieve Coreference Resolution for {I}ndonesian},
author = {Artari, Valentina Kania Prameswara and Mahendra, Rahmad and Jiwanggi, Meganingrum Arista and Anggraito, Adityo and Budi, Indra},
year = 2021,
month = sep,
booktitle = {Proceedings of the International Conference on Recent Advances in Natural Language Processing (RANLP 2021)},
publisher = {INCOMA Ltd.},
address = {Held Online},
pages = {79--85},
url = {https://aclanthology.org/2021.ranlp-1.10},
abstract = {Coreference resolution is an NLP task to find out whether the set of referring expressions belong to the same concept in discourse. A multi-pass sieve is a deterministic coreference model that implements several layers of sieves, where each sieve takes a pair of correlated mentions from a collection of non-coherent mentions. The multi-pass sieve is based on the principle of high precision, followed by increased recall in each sieve. In this work, we examine the portability of the multi-pass sieve coreference resolution model to the Indonesian language. We conduct the experiment on 201 Wikipedia documents and the multi-pass sieve system yields 72.74{\%} of MUC F-measure and 52.18{\%} of BCUBED F-measure.}
}
License
MIT
Homepage
https://github.com/valentinakania/indocoref/
NusaCatalogue
For easy indexing and metadata: https://indonlp.github.io/nusa-catalogue