SPACCC_SPLIT: Spanish Clinical Case Corpus - Sentence Splitting
Digital Object Identifier (DOI) and access to dataset files
This repository contains the sentence splitting annotations in the Spanish Clinical Case Corpus. 20% of the corpus was annotated manually by two annotators, while the remaining 80% was annotated automatically with the Spanish Clinical Case Corpus Part-of-Speech Tagger (SPACCC_POS-TAGGER, https://github.com/PlanTL/SPACCC_POS-TAGGER), with an implemented version of the FreeLing3.1 tool, which mimics the criteria marked by the two human annotators.
The corpus has 64,865 sentences, 353,144 words and 18,281 different lemmata. The ratio of words per sentence is 5.44.
corpus/ Original, development, validation and automatically annotated corpus, both in tabular format and BRAT standoff format. guidelines/ Annotation guidelines. IAA/ Inter-annotator agreement report, along with the data and the scripts used to calculate it. script/ Script to convert FreeLing3.1 tabular output format into BRAT standoff format.
The SPACCC corpus was created after collecting 1,000 clinical cases from SciELO (Scientific Electronic Library Online), an electronic library that gathers electronic publications of complete full text articles from scientific journals of Latin America, South Africa and Spain (http://www.scielo.org).
A clinician classified those cases into those that were similar to real clinical texts in terms of structure and content and those that were not suitable for this task. Figure legends were automatically removed and, in case multiple clinical cases were listed, these were split into single clinical cases.
Annotations were carried out by means of the Spanish Clinical Case Corpus Part-of-Speech Tagger based on FreeLing3.1 (SPACCC_POS-TAGGER, https://github.com/PlanTL/SPACCC_POS-TAGGER).
Annotations created in SPACCC_SPLIT are provided in BRAT standoff format; i.e. the annotations are stored separately
.ann file) from the document text (a
These two files are associated by their base name; their file name without suffix is the same, for example, the file
es-S0004-06142005000200009-1.ann contains the annotations for the file
See http://brat.nlplab.org/standoff.html for further details on the brat standoff format.
This annotation format is produced by running an script that converts the output of SPACC_POS-TAGGER, a CoNLL-like column format, where columns are:
FORM: word form.
LEMMA: word lemma.
TAG: complete POS tag.
PROBABILITY: probability of the chosen tag.
.ann file only Sentence Boundary Simbols are marked and they are labeled with FreeLing's tag Fp.
The quality of the annotations at the level of sentence splitting carried out by FreeLing prior its adaptation to the
clinical corpus, was measured with the development corpus: 100 randomly chosen texts (10% of the whole corpus).
99.37% of this corpus was successfully annotated. Discrepancies were generated due to abbreviations in final sentence position ending in point (e.g: 'stage Ib._Posteriormente', 'dcha._La' 'mmol / L._Recibía') and segmentations of acronyms that include points in several tokens (e.g.: i.m).
The annotation guidelines describe the criteria that have been followed to annotate the corpus, along with illustrative examples. They describe FreeLing default resources, the criteria that have been followed in the manual annotation and the implementations that solve these criteria in automatic annotation. The guidelines also compare the criteria followed in this project with those followed by the Apache CTAKES NLP system (http://ctakes.apache.org/) and the the GENIA Sentence Splitter (http://www.nactem.ac.uk/y-matsu/geniass/), as well as the criteria followed in the development of the GENIA corpus (https://github.com/spyysalo/genia-pos).
Guidelines have been written and developed in Spanish and are only available in Spanish.
The following three tables show the interagreement results measured on both the development and the validation corpus. See the inter-annotator agreement report (Informe_interagreement_CNIO_PlanTL_SEAD.pdf) included in folder
IAA in this repository for further details. Note that the required minimum level was 99%.
|A1 vs A2||99,79%|
|A1 vs FL||99,37%|
|A2 vs FL||99,58%|
Table 1: Interagreement betwen the two human annotators and SPACCC_POS_TAGGER on the development corpus.
|GS vs FL||99,37%|
Table 2: Interagreement between the gold standard corpus and SPACCC_POS-TAGGER on the development corpus.
|GS vs FL||99,52%|
Table 3: Interagreement between the gold standard corpus and SPACCC_POS-TAGGER on the validation corpus.
Montserrat Marimon (email@example.com)
This work is licensed under a Creative Commons Attribution 4.0 International License.
Copyright (c) 2018 Secretaría de Estado para el Avance Digital (SEAD)