README.md

metadata

dataset_info:
  features:
    - name: id
      dtype: string
    - name: long_text
      dtype: string
    - name: short_text
      dtype: string
    - name: style
      list: string
    - name: labels_text
      list: string
    - name: labels
      list: int64
    - name: licenses
      struct:
        - name: id
          dtype: string
        - name: labels
          dtype: string
        - name: labels_text
          dtype: string
        - name: long_text
          dtype: string
        - name: short_text
          dtype: string
        - name: style
          dtype: string
        - name: terms_of_use
          dtype: string
  splits:
    - name: train
      num_bytes: 6567091
      num_examples: 1811
    - name: valid
      num_bytes: 1332559
      num_examples: 376
    - name: test
      num_bytes: 1332559
      num_examples: 376
  download_size: 5195698
  dataset_size: 9232209
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: valid
        path: data/valid-*
      - split: test
        path: data/test-*
license: cc-by-4.0
language:
  - en
pretty_name: National Gallery Artwork Style Classification
size_categories:
  - 1K<n<10K
task_categories:
  - text-classification
tags:
  - art
  - art-history
  - glam
  - lam
  - ai4lam

National Gallery Artwork Style Classification

This dataset is one of a collection of datasets, derived from the National Gallery Elasticsearch endpoint (https://data.ng.ac.uk/es/public/_search). Each row comprises the long- and short-text descriptions, found on the Gallery's website (e.g. Sunflowers - Vincent van Gogh). From this textual-description alone, the model must predict the one or more style classifications, designated by art historical experts at the National Gallery.

The output from the Elasticsearch endpoint is first flattened and unwanted keys discarded using Proteus. Rows with null or empty short- or long-texts are removed. A style label must have a minimum frequency of 5 occurences across the collection to be kept. After filtering out low-frequency labels, rows with a null or empty style field are removed from the dataset.

The following labels are included for classification:

Style	Frequency
Baroque	900	Impressionist	125	Stuart	19
High Renaissance	623	Barbizon	76	Pointillist	12
Early Renaissance	534	Georgian	75	International Gothic	10
Gothic	336	Romantic	75	Symbolist	10
Neoclassical	240	Post-Impressionist	35	Nabi	9
Academic	215	Caravaggist	33	Regency	8
Rococo	162	Mannerist	27	Modernist	6
Realist	134	Victorian	26

Label Mapping

The following code can be used to convert predictions back into text labels:

import numpy as np

ordered_label_text = np.array([
 'Academic', 'Barbizon', 'Baroque', 'Caravaggist', 'Early Renaissance', 'Georgian', 'Gothic', 'High Renaissance',
 'Impressionist', 'International Gothic', 'Mannerist', 'Modernist', 'Nabi', 'Neoclassical', 'Pointillist',
 'Post-Impressionist', 'Realist', 'Regency', 'Rococo', 'Romantic', 'Stuart', 'Symbolist', 'Victorian'
])

# Single example
label = np.array([0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0])
ordered_text_labels[label == 1]

# Apply to dataframe column
df["predictions"].apply(lambda pred: ordered_label_text[pred == 1])

Limitations and Biases

Reflecting the collecting priorities of the National Gallery itself, this dataset exclusively represents European art, specifically that of Western Europe. The collection spans from the 13th to the early 20th century, therefore excluding many entire movements in more recent art-history (e.g. dadaism, surrealism) and those before approximately 1200 (e.g. Early Christian, Pre-Romanesque). Furthermore, the works in the collection are predominantly by well-established artists in the Western art historical canon.

The distribution of labels is also heavily imbalanced: 14 labels only appear once across the collection (Wan-Li, Medieval, Cubist, Fauve, Pre-Raphaelite, Expressionist, Plantagenet, Viennese Secessionist, Die Brücke, Surrealist, Ottoman, Edwardian, Classical, Aesthetic Movement); and a further 4 are also excluded due to low frequency (Renaissance, Antique, Nazarene, Tudor). As these labels were assigned over an extended period and among many individuals, different degrees of granularity were adopted, creating groups of labels that form hierarchies of generality (e.g. Renaissance ⊇ {High Renaissance, Early Renaissance}). A future version of this dataset will try to make use of these hierarchies.

Users should be aware that this dataset encodes particular art-historical perspectives that:

• Reflect Western/European art historical frameworks
• May perpetuate canonical hierarchies in art history
• Represent institutional collecting practices of a single organisation

License & Copyright Statement

To encourage the use and reuse of the National Gallery's collection data, they are released under the following dedications and licences:

Structured data (as opposed to narrative texts) are released under a Creative Commons Zero dedication (CC0):
https://creativecommons.org/publicdomain/zero/1.0/.

Descriptions, notes and all other narrative text content are licensed under a Creative Commons Attribution 4.0 licence (CC BY):
https://creativecommons.org/licenses/by/4.0/.

Images are licensed under a Creative Commons Attribution-NonCommercial-NoDerivs 4.0 licence (CC BY-NC-ND):
https://creativecommons.org/licenses/by-nc-nd/4.0/.