metadata

license: cc-by-nc-sa-4.0
language:
  - es
pretty_name: 'EsCoLA: Spanish Corpus of Linguistic Acceptability'
configs:
  - config_name: default
    data_files:
      - split: train
        path: EsCoLA_InDomainTrain.tsv
      - split: validation
        path: EsCoLA_InDomainDev.tsv

Introduction

The Spanish Corpus of Linguistic Acceptability (EsCoLA) includes 11,174 sentences taken from linguistic literature with a binary annotation made by the original authors themselves. The work is inspired by CoLA: https://nyu-mll.github.io/CoLA/#

Paper

Núria Bel, Marta Punsola, Valle Ruiz-Fernández, 2024, EsCoLA: Spanish Corpus of Linguistic Acceptability. Joint International Conference on Computational Linguistics, Language Resources and Evaluation LREC-COLING 2024. Torino. Italy.

Download

The corpus has a CC-BY 4.0 license. Download EsCoLA inDomain train and dev datasets, plus human annotation, from https://github.com/nuriabel/LUTEST/ For EsCoLA outDomain dataset and EsCoLA inDomain test data, please contact [email protected].

Data format

EsCoLA dataset is split into two subsets: an in-domain subset (InDomain) with 10,567 sentences, and an out-of-domain subset (OutDomain) with 607 sentences. The in-domain subset has been split into train/dev/test sections:

train: 8454 sentences

dev: 1053 sentences

test: 1060 sentences

And the out-of-domain subset is split into dev/test sections. The test sets are not made public.

For the in-domain subset, each line in the .tsv files consists of 11 tab-separated columns:

Column 1: a unique ID Column 2: the source of the sentence Column 3: the acceptability judgment label from the source (0=unacceptable, 1=acceptable) Column 4: the source's annotation (* for the unacceptable sentences) Columns 5, 6 and 7: the human annotations Column 8: the human annotations' median Column 9: the sentence Column 10: the category of the linguistic phenomenon the sentence is an example of Column 11: the split to which the sentence belongs

For the out-of-domain, each line in the .tsv file consists of 6 tab-separated columns:

Column 1: a unique ID Column 2: the source of the sentence Column 3: the acceptability judgment label from the source (0=unacceptable, 1=acceptable) Column 4: the source's annotation (* for the unacceptable sentences) Column 5: the sentence Column 6: the category of the linguistic phenomenon the sentence is an example of

Processing

During the gathering of the data and their processing, some sentences from the source documents may have been omitted or altered. We discarded examples marking dubious acceptability with "?" or other signs, but those examples that included acceptability alternations were taken by creating the two versions: the acceptable and the unacceptable sentence. Finally, the examples that were not full sentences, that is, that contain no main verb, were manually edited to add a neutral verb to convert them into sentences, while keeping the acceptability value.

Sources

InDomain: Demonte and Bosque (1999) OutDomain: RAE (2009), Palencia and Aragonés (2007) Díaz and Yagüe (2019)

Annotation

The dataset has been manually annotated with 14 linguistic phenomena.

Simple
Predicative
Adjuncts
Argument types
Argument alternation
Binding pronouns
Wh-phenomena
Complement clauses
Modal, negation, periphrasis and auxiliaries
Infinitive embedded VPs and referential phenomena
Complex NPs and APs
S-Syntax
Determiners, quantifiers, comparative and superlative constructions
Spanish phenomena The Spanish phenomena have been classified into 6 categories: 14.1. Agreement in nominal constructions 14.2. Subjunctive mode and tense 14.3. Spurious preposition for completive clauses ('dequeismo') 14.4. Subject ellipsis 14.5. Pronominal cliticization 14.6. Ser/estar copula selection

Citation

Please, if you use the dataset cite the following papers:

Alex Warstadt, Amanpreet Singh, and Samuel R. Bowman. 2018. Neural network acceptability judgments. arXiv preprint arXiv:1805.12471.

Núria Bel, Marta Punsola, Valle Ruiz-Fernández, 2024, EsCoLA: Spanish Corpus of Linguistic Acceptability. Proceedings of the Joint International Conference on Computational Linguistics, Language Resources and Evaluation LREC-COLING 2024. Torino. Italy.

Disclaimer

The dataset has been made by copying the examples from published works that are protected by copyright. According to Spanish law, we have respected the copyright because the number of elements taken represent less than a 10% of the whole work, and the number of items copied is justified by the aims of research.