---
license: cc-by-nc-sa-4.0
language:
- be
tags:
- morpheme segmentation
- tokenization
- morphemes
- Belarusian
pretty_name: Slounik-Wordforms
size_categories:
- 100K<n<1M
---

# Slounik-Wordform (Belarusian word-form morpheme segmentation dataset)

This dataset contains morpheme segmentations for Belarusian word forms. It extends the [Slounik](https://huggingface.co/datasets/ruscorpora/morphodict-bel) dataset, a Belarusian morpheme segmentation resource based on the *School Morpheme Dictionary of the Belarusian Language* (Mormysh et al., 2005).

The dataset follows a *surface segmentation* approach: morphemes are segmented as they appear in the word form, without restoring underlying or historical forms. For example, morphophonemic alternations are preserved in their surface form rather than normalized.

Each morpheme is labeled with its type. The main morpheme types are:

* `PREF` — prefix
* `ROOT` — root
* `SUFF` — suffix
* `END` — inflectional ending
* `LINK` — linking vowel / interfix
* `HYPN` — hyphen

In total, the dataset contains **332,497** unique morpheme segmentations of Belarusian word forms:

| POS        | Number of forms |
| ---------- | --------------: |
| Verbs      |         150,543 |
| Nouns      |          86,442 |
| Adjectives |          93,037 |
| Adverbs    |           2,475 |
| **Total**  |     **332,497** |

The dataset contains **8,186** unique morphemes. The average number of morphemes per segmentation is **3.86**.

The segmentations are stored line by line. Each line consists of the word form, a tabulation symbol, and the segmentation. Morphemes in the segmentations are separated with `/`, and each morpheme has the form `MORPHEME:TYPE`.

Data example:

```text
зябну	зяб:ROOT/н:SUFF/у:END
тэматычнаму	тэм:ROOT/атыч:SUFF/н:SUFF/аму:END
належнымі	належ:ROOT/н:SUFF/ымі:END
адпускнікоў	ад:PREF/пуск:ROOT/н:SUFF/ік:SUFF/оў:END
пераскварваеш	пера:PREF/сквар:ROOT/ва:SUFF/еш:END
маркоціў	маркоц:ROOT/і:SUFF/ў:SUFF
зменшылі	з:PREF/мен:ROOT/ш:SUFF/ы:SUFF/л:SUFF/і:END
нагружаць	на:PREF/груж:ROOT/а:SUFF/ць:SUFF
```

## Data Preparation

The dataset was prepared by the [Russian National Corpus](https://ruscorpora.ru/en) team.

The starting point was the [Slounik](https://huggingface.co/datasets/ruscorpora/morphodict-bel) dataset, which contains 35,219 Belarusian word forms annotated according to a surface morpheme segmentation scheme. The original Slounik dataset includes four parts of speech: nouns, adjectives, verbs, and adverbs. Adverbs are uninflected, while nouns, adjectives, and verbs have rich inflectional paradigms.

To generate word-form segmentations, we used the [Grammatical Database of the Belarusian Language](https://github.com/Belarus/GrammarDB) (Koshchanka et al., 2021). For each lemma, we derived a stem by removing the inflectional ending and, for verbs, the infinitive suffix *-ць*. We then compared the resulting stem with the target word form and identified the string difference between them.

If the transformation consisted of simple right-side concatenation, the appended material was segmented according to POS-specific inflectional and derivational patterns. If the difference involved internal variation, the change was checked against a predefined list of morphophonemic alternations. For example:

```text
сел:ROOT/ян:SUFF/ін:SUFF → сял:ROOT/ян:SUFF/е:END
абавяз:ROOT/ак:SUFF → абавяз:ROOT/к:SUFF/у:END
```

The appended strings were then verified against POS-specific patterns and segmented into morphemes.

## Validation

The resulting dictionary underwent two rounds of semi-manual selective validation by a native Belarusian speaker with an advanced degree in linguistics. Based on the validation results, the lists of alternations and segmentation patterns were refined. All identified systematic errors were corrected.

One known limitation concerns rare cases of morphological homonymy in verb forms, where the same surface form may correspond to two different segmentations. For example:

```text
над:PREF/кус:ROOT/і:SUFF/це:POST
над:PREF/кус:ROOT/іце:END
```

or:

```text
рас:PREF/стро:ROOT/і:SUFF/ць:SUFF
рас:PREF/стро:ROOT/іць:END
```

In such cases, one of the possible segmentations was assigned at random. We estimate that these cases account for less than **0.1%** of the total dataset.

## Data Usage

This dataset can be used for training and evaluating automatic morpheme segmentation systems for Belarusian word forms. It may also be useful for experiments on morphology-aware tokenization, morphology-aware language modeling, lemmatization, word-form generation, and processing of morphologically rich Slavic languages.

The source Slounik dataset was used in the ACL 2025 paper [BERT-like Models for Slavic Morpheme Segmentation](https://aclanthology.org/2025.acl-long.337/) (Morozov et al., 2025).

## License

The Slounik-Wordform dataset is publicly available under the **CC-BY-NC-SA 4.0** license.

Since the dataset is derived from the Slounik dataset and ultimately from the *School Morpheme Dictionary of the Belarusian Language* (Mormysh et al., 2005), any use of the dataset must include a mandatory mention of the authors of the original dictionary and a link to the BSPU repository: [A](http://elib.bspu.by/handle/doc/30574), [B](http://elib.bspu.by/handle/doc/30575), [C](http://elib.bspu.by/handle/doc/30576), [D](http://elib.bspu.by/handle/doc/30577), [E](http://elib.bspu.by/handle/doc/30578), [F](http://elib.bspu.by/handle/doc/30579), [G](http://elib.bspu.by/handle/doc/30580)

The dataset also relies on the Grammatical Database of the Belarusian Language. Please acknowledge this resource when using the generated word-form segmentations.

## How to cite

If you use this dataset in scientific work, please cite the source Slounik dataset paper:

**Dmitry Morozov, Lizaveta Astapenka, Anna Glazkova, Timur Garipov, and Olga Lyashevskaya.** 2025. [BERT-like Models for Slavic Morpheme Segmentation](https://aclanthology.org/2025.acl-long.337/). In *Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)*, pages 6795–6815, Vienna, Austria. Association for Computational Linguistics.