metadata

license: cc-by-4.0
dataset_info:
  features:
    - name: idiom
      dtype: string
    - name: is_pie
      dtype: bool
    - name: tokens
      sequence: string
    - name: ner_tags
      sequence:
        class_label:
          names:
            '0': O
            '1': B-PIE
            '2': I-PIE
  splits:
    - name: train
      num_bytes: 82950018
      num_examples: 46090
    - name: validation
      num_bytes: 10420303
      num_examples: 5761
    - name: test
      num_bytes: 10376839
      num_examples: 5762
  download_size: 19258913
  dataset_size: 103747160
task_categories:
  - token-classification
language:
  - en
tags:
  - PIE
  - idioms
size_categories:
  - 10K<n<100K
pretty_name: Corpus of potentially idiomatic expressions (PIEs)

Dataset Card for PIEs corpus

Dataset Summary

This corpus is a collection of 57170 potentially idiomatic expressions (PIEs) based on the British National Corpus, prepaired for NER task. Each of the objects is comes with a contextual set of tokens, BIO tags and boolean label. The data sources are:

Detailed data preparation pipeline can be found here

Supported Tasks and Leaderboards

Token classification (NER)

Languages

English

Dataset Structure

Data Instances

For each instance there is a string with target idiom, tokenized by word text with context of idiom usage, corresponded BIO tags and boolean label is_pie. This tag determines whether or not a collocation is considered an idiom in a given context. For a PIE dataset the choice was determined by the original PIE_label. For MAGPIE a threshold of 0.75 confidence coefficient was chosen.

An example from the train set looks like the following:

{'idiom': "go public"
'is_pie': True
'tokens': [ "Private", "dealers", "in", "the", "States", "go", "public" ]
'ner_tags': [ 0, 0, 0, 0, 0, 1, 2 ]
}

Where NER tags is {0: 'O', 1: 'B-PIE', 2: 'I-PIE'}

Data Fields

idiom: a string containg original PIE
is_pie: a boolean label determining whether a PIE can be considered an idiom in a given context
tokens: sequence of word tkenized string with PIE usage context
ner_tags: corresponded BIO tags for word tokens

Data Splits

The SNLI dataset has 3 splits: train, validation, and test.

Dataset Split	Number of Instances in Split
Train	45,736
Validation	5,717
Test	5,717

Dataset Creation

Source Data

Initial Data Collection and Normalization

Additional Information

Licensing Information

Corpus and it's sources are licensed under a Creative Commons Attribution-ShareAlike 4.0 International License.

Citation Information

PIE Corpus (Haagsma, H. (Creator), Bos, J. (Contributor), Plank, B. (Contributor), University of Groningen.)
MAGPIE: A Large Corpus of Potentially Idiomatic Expressions (Haagsma et al., LREC 2020)