Dataset Card for morphoseg-en

Dataset Summary

• Owner: browndw
• Name: morphoseg-en
• Task: English word-to-morpheme segmentation
• Source: Processed from the Wiktionary-derived corpus released via Zenodo (DOI: 10.5281/zenodo.5172856)
• Format: JSONL with fields word (string) and segments (list of strings)

Supported Tasks and Leaderboards

• Morphological Segmentation: Sequence-to-sequence prediction of morpheme boundaries and affix strings.
• No official leaderboard is associated with this dataset at the moment.

Languages

• English (en). Foreign-language affixes may appear in etymological compounds but the primary target language is English.

Dataset Structure

• Total examples: 275,836 unique word/segments pairs with at least two morphemes.
• Splits:
  • train: 220,668 examples (80%)
  • val: 27,583 examples (10%)
  • test: 27,585 examples (10%)
• Each example contains the raw word form and an ordered list of morpheme strings (including prefixes/suffixes such as "non-", "-ation").

Data Processing

• Stream-parsed the original wiki_morph.json using ijson to avoid loading the entire file in memory.
• Extracted (word, segments) pairs from entries with non-empty Morphemes arrays.
• Filtered out examples with fewer than two morphemes (--min-segments 2).
• Deduplicated by exact (word, tuple(segments)) to remove repeated senses.
• Randomly shuffled and split the resulting examples into train/validation/test partitions with seed 42.

Licensing and Usage

• The original dataset at DOI 10.5281/zenodo.5172856 should be consulted for licensing terms (the upstream release is CC BY-SA 3.0 per Zenodo metadata). Downstream users must comply with the original license conditions, including attribution and share-alike provisions when applicable.
• This processed dataset inherits the same licensing constraints. When redistributing or deploying models trained on morphoseg-en, include attribution to both the original Zenodo release and this repository.

Intended Uses

• Training and evaluating morphological segmentation models (e.g., byte/character-level Transformers, sequence taggers).
• Linguistic analysis of morpheme distributions in English lexicon data.

Limitations

• Derived from Wiktionary entries; coverage favors lexemes with richer dictionary documentation and may exhibit inconsistencies in affix notation.
• Homographs with divergent segmentations may exist, but sense-level context is stripped, so models see only surface forms.
• Some entries retain historical or etymological morpheme analyses rather than synchronic segmentations.

Citation

If you use morphoseg-en in your work, please cite the Zenodo source and this processed release. Example BibTeX:

@dataset{morphoseg_en,
  author       = {Brown, David West},
  title        = {morphoseg-en: English Morpheme Segmentation Dataset},
  year         = {2025},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/browndw/morphoseg-en}}
}

And the original source:

@dataset{prazak_2021_morphology,
  author       = {Prazak, David and others},
  title        = {Morphological segmentation dataset from Wiktionary},
  year         = {2021},
  publisher    = {Zenodo},
  doi          = {10.5281/zenodo.5172856}
}

Contact

For questions or issues, file an issue in the GitHub repository associated with this project or reach out to browndw on Hugging Face.