README.md

metadata

language:
  - en
license: cc-by-4.0
dataset_info:
  features:
    - name: task
      dtype: string
    - name: subtask
      dtype: string
    - name: similarity
      dtype: float64
    - name: speaker_id
      dtype: int64
    - name: pair_id
      dtype: int64
    - name: audio_a
      dtype:
        audio:
          sampling_rate: 16000
    - name: audio_b
      dtype:
        audio:
          sampling_rate: 16000
    - name: sentence_a
      dtype: string
    - name: sentence_b
      dtype: string
  splits:
    - name: test
      num_bytes: 1713645707.328
      num_examples: 2552
  download_size: 1575109909
  dataset_size: 1713645707.328
configs:
  - config_name: default
    data_files:
      - split: test
        path: data/test-*

SpokenSTS Dataset

Spoken versions of the Semantic Textual Similarity dataset for testing semantic sentence level embeddings. Contains thousands of sentence pairs annotated by humans for semantic similarity. The spoken sentences can be used in sentence embedding models to test whether your model learns to capture sentence semantics.

Disclaimer

This distribution is not official. This subset only contains sentences from 4 human voices. Synthesized voices are excluded.

Dataset structure

• There are five tasks: STS12 ~ STS16.
• Each tasks has couple of subtasks. Each subtask as few dozens of unique sentence pairs (numbers in parenthesis).
  • STS12: MSRpar (38), MSRvid (38), SMTeuroparl (23), OnWN (43), SMTnews (20)
  • STS13: FNWN (10), headlines (38), OnWN (30)
  • STS14: deft-forum (23), deft-news (15), headlines (38), images (38), OnWN (32), tweet-news (38)
  • STS15: answers-forums (19), answers-students (38), belief (19), headlines (38), images (38)
  • STS16: answer-answer (13), headlines (13), plagiarism (12), postediting (13), question-question (11)
• Total 638 unique sentence pairs exist. Note that pair_id is only unique within the subtask.
• For each sentence pair, there are utterances from 4 speakers (speaker_id 1 ~ 4), total 638x4=2552 rows in the dataset.
• Sentence pair similarity ranges from minimum 0.0 to maximum 5.0.
• Refer Table 1 of SpokenSTS paper for more details on tasks & subtasks.
• Audio is resampled into 16kHz.

References

• Original dataset and detailed metadata can be found at https://doi.org/10.17026/dans-z48-3ev6
• Codebase of SpokenSTS can be found at https://github.com/DannyMerkx/speech2image/tree/Interspeech21
• If you use the dataset, please cite:
  • (Original STS database) Eneko Agirre, Carmen Banea, Daniel Cer, Mona Diab, Aitor Gonzalez-Agirre, Rada Mihalcea, German Rigau, and Janyce Wiebe. Semeval-2016 task 1: Semantic textual similarity, monolingual and cross-lingual evaluation. In SemEval, pages 497–511. ACL, 2016.
  • (SpokenSTS) Danny Merkx, Stefan L. Frank and Mirjam Ernestus (2021). Semantic sentence similarity: size does not always matter. In Interspeech 2021 - 22nd Annual Conference of the International Speech Communication Association. pp. pp. 4393-4397