README.md

---
license: cc-by-nc-sa-4.0
task_categories:
- token-classification
language:
- ru
tags:
- semantics
- Russian
- RNC
pretty_name: RNC Semantics
---

# RNC Semantics

## Dataset Summary

RNC Semantics is a collection of Russian semantic annotation resources developed for experiments in automated semantic markup of corpus data. The repository contains:

- a compact semantic tagset designed for Russian corpus annotation;
- annotated datasets for semantic tagging and word sense disambiguation.

The annotation scheme is based on **atomic semantic tags**. Instead of assigning hierarchically ordered labels, in our scheme semantic specificity is represented compositionally through combinations of simpler tags. For example, an evaluative positive word can be annotated as `ev` + `posit`, and a word denoting a means of transportation can be annotated as `tool` + `transp`.

The dataset was created in the context of the paper *Rethinking Semantic Markup in the Russian National Corpus: An Annotation Schema and ML Baselines*.

## Intended Uses

This resource is intended for:

- semantic tagging of Russian corpus data;
- word sense disambiguation for polysemous Russian lemmas;
- evaluation of contextualized encoder models and LLM-based semantic tagging.

The resource is not intended to provide full lexical sense definitions in the style of WordNet. It uses compact semantic tags and tag combinations rather than fine-grained sense inventories.

## Annotation Scheme

The proposed scheme contains **83 atomic semantic tags**. Tags are designed to be combined when a single lexical item expresses several semantic dimensions. This makes the scheme compact while preserving expressive power.

The complete tag inventory is distributed separately as `RNC-semantics-scheme.csv`. The README intentionally includes only a few examples; please use the CSV file as the authoritative source for the full list of tags, descriptions, and examples.

The CSV file has the following columns:

| Column | Description |
|---|---|
| `tag` | Short tag name used in the annotation. |
| `description` | Explanation of the semantic category. |
| `examples` | Illustrative English examples. |

### Tag Examples

| Tag | Description | Examples |
|---|---|---|
| `color` | Denotes color properties of objects or phenomena, including direct color terms and words that indirectly describe color characteristics. | yellow, black |
| `food` | Refers to food and drink items, food sets, and tastes. | soup, cheese |
| `plant` | Refers to any representatives of the plant kingdom — trees, flowers, shrubs, grasses, fungi, algae, their parts, and general group names. | tree, leaf, cereals |
| `speed` | Describes the rapidity or slowness of movement, change, or progression of processes. | fast, tempo, pace |
| `hum` | Refers to individual persons, groups of people, or parts of the human body, including names, professions, family relations, ethnic groups, and body parts when clearly referring to a human. | mother, John, russians |
| `abstr` | Refers to abstract concepts denoting phenomena, ideas, feelings, qualities, processes, or states that have no physical form and cannot be perceived by the senses. | love, justice |
| `ev` | Expresses subjective judgment, opinion, or attitude toward an object, action, or quality. | beautiful, disgusting |
| `transp` | Refers to all types of vehicles, their classes, and parts of vehicles. | car, wheel, bicycle |

### Compositional Annotation

The scheme is compositional. A token may receive more than one semantic tag when a single word contributes several semantic properties in context. Examples:

| Word / meaning | Possible tag combination |
|---|---|
| *good* as positive evaluation | `ev` + `posit` |
| *boat* as a means of transport | `tool` + `transp` |
| *drinking water* | `concr` + `food` + `stuff` |
| *fish* as food | `concr` + `food` |
| *fish* as an animal | `concr` + `class` + `animal` |

## Dataset Subsets

| Subset | Annotation scenario | Source / method | Size | Tagged examples | Unique labeled lemma+POS combinations |
|---|---|---|---:|---:|---:|
| RNC-Manual | Expert annotation of continuous text | Manual token-by-token annotation by RNC experts | ~6k tokens | ~2.7k tokens | 1,451 |
| RNC-Crowdsource | Word sense disambiguation | Crowdsourced annotation of 16 frequent polysemous lemmas, with majority voting | ~115k tokens in contexts | ~5.2k disambiguated examples | 16 |
| RNC-SynTagRus | Distant supervision | Projection from the RNC semantic lexicon to unambiguous lexical items in [SynTagRus](https://github.com/UniversalDependencies/UD_Russian-SynTagRus/) | 1,517k tokens | 248k tokens | 18,579 |
| RNC-CoBaLD | Cross-framework conversion | Rule-based conversion from [CoBaLD](https://github.com/CobaldAnnotation/CobaldRus) semantic classes to the proposed scheme | 443k tokens | 54k tokens | 5,534 |

## Annotation Format

The annotated datasets are distributed in **CoNLL-U** format. CoNLL-U is a plain-text, tab-separated format in which each token line contains 10 fields:

```text
ID    FORM    LEMMA    UPOS    XPOS    FEATS    HEAD    DEPREL    DEPS    MISC
```

Sentence-level metadata is stored in comment lines beginning with `#`, for example `# sent_id`, `# genre`, and `# text`. Sentences are separated by a blank line.

Semantic annotation is stored in the **MISC** column. The main semantic attribute is `sem`, whose value is a list of tags from `RNC-semantics-scheme.csv`:

```text
sem=['concr', 'device', 'tool']
```

If a token has several semantic tags, they are listed together inside the `sem` value. If a token has no semantic annotation, the `sem` attribute is omitted. Other MISC attributes are preserved and separated with `|`, as in standard CoNLL-U-style notation:

```text
SpaceAfter=No|sem=['sound']|hls
```

The auxiliary flag `hls`, when present, is not a semantic tag and should not be treated as part of the tag inventory. The authoritative list of semantic tags is provided in `RNC-semantics-scheme.csv`.

### Example

```conllu
# sent_id = GramEval2020-test-fiction-2
# genre = fiction
# text = Может быть, потому, что музыкальный автомат грохотал здесь не так оглушительно, как везде.
1	Может	мочь	VERB	_	Aspect=Imp|Mood=Ind|Number=Sing|Person=3|Tense=Pres|Transit=Intr|VerbForm=Fin|Voice=Act	4	parataxis:discourse	_	sem=['abstr']|hls
2	быть	быть	VERB	_	Transit=Intr|VerbForm=Inf|Voice=Act	1	xcomp	_	SpaceAfter=No
3	,	,	PUNCT	_	_	1	punct	_	_
4	потому	потому	ADVPRO	_	Degree=Pos|PronType=Dem	0	root	_	SpaceAfter=No
5	,	,	PUNCT	_	_	9	punct	_	_
6	что	что	SCONJ	_	_	9	mark	_	_
7	музыкальный	музыкальный	ADJ	_	Case=Nom|Degree=Pos|Gender=Masc|Number=Sing	8	amod	_	sem=['mus', 'sound']|hls
8	автомат	автомат	NOUN	_	Animacy=Inan|Case=Nom|Gender=Masc|Number=Sing	9	nsubj	_	sem=['concr', 'device', 'tool']|hls
9	грохотал	грохотать	VERB	_	Aspect=Imp|Gender=Masc|Mood=Ind|Number=Sing|Tense=Past|Transit=Intr|VerbForm=Fin|Voice=Act	4	advcl	_	sem=['sound']|hls
10	здесь	здесь	ADVPRO	_	Degree=Pos|PronType=Dem	9	advmod	_	sem=['place']|hls
11	не	не	PART	_	Polarity=Neg	12	advmod	_	_
12	так	так	ADVPRO	_	Degree=Pos|PronType=Dem	13	advmod	_	sem=['degr']|hls
13	оглушительно	оглушительно	ADV	_	Degree=Pos	9	advmod	_	SpaceAfter=No|sem=['sound']|hls
14	,	,	PUNCT	_	_	16	punct	_	_
15	как	как	SCONJ	_	_	16	mark	_	_
16	везде	везде	ADVPRO	_	Degree=Pos|PronType=Tot	12	advcl	_	SpaceAfter=No|sem=['place']|hls
17	.	.	PUNCT	_	_	4	punct	_	_
```

## Dataset Creation

### RNC-Manual

RNC-Manual was created to test semantic annotation in continuous running text. A subset of texts from the Russian National Corpus was annotated exhaustively by linguistic experts. Non-semantic tokens such as punctuation and foreign words were excluded from semantic evaluation.

### RNC-Crowdsource

RNC-Crowdsource targets word sense disambiguation. The dataset contains examples for 16 frequent polysemous Russian lemmas. Each example was annotated independently by three volunteers. The final label was assigned by majority vote and retained only when at least two annotators selected the same sense.

### RNC-SynTagRus

RNC-SynTagRus was created through distant supervision. Semantic tags from the RNC semantic lexicon were projected onto SynTagRus tokens, but only for lexical items associated with a single unique semantic tagset. This filtering reduces ambiguity-related noise while providing a larger training set.

### RNC-CoBaLD

RNC-CoBaLD was created by converting semantic annotations from the Russian CoBaLD corpus into the proposed RNC semantic scheme. Because CoBaLD assigns one semantic class per token while the proposed scheme allows tag combinations, converted annotations are expected to be high-precision but may be less complete than native annotations.

## Known Limitations

- The scheme is intentionally compact and does not attempt to encode full lexical sense definitions.
- Some semantic distinctions require combinations of tags rather than a single label.
- Automatically derived subsets may contain noise or incomplete tagsets.
- RNC-CoBaLD conversion is necessarily partial because the source and target schemes differ structurally.
- RNC-SynTagRus excludes polysemous items by design and therefore does not represent the full ambiguity of Russian vocabulary.
- Small or minority senses remain challenging for automatic models and may be underrepresented.

## Citation

If you use this dataset, please cite the accompanying paper:

```bibtex
@inproceedings{smal_morozov_openko_rnc_semantics,
  title = {Rethinking Semantic Markup in the Russian National Corpus: An Annotation Schema and ML Baselines},
  author = {Smal, Ivan and Morozov, Dmitry and Openko, Vitaliya},
  year = {2026},
  note = {Forthcoming}
}
```

## License

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