ChangeIsKey/denotational-roberta-classifier

Denotational Type Classifier

This model classifies the denotational relation between two word senses, represented as a pair of definitions. It is based on roberta-large with a sequence-classification head and is trained for lexical-semantic relation classification.

The model predicts one of the following relation types:

• generalization
• specialization
• metaphor
• metonymy
• homonymy
• antonymy

The first five labels correspond to the denotational relation types evaluated in SenseRel. antonymy is included because it is present in the fine-tuning data, although it is not part of the expert-annotated SenseRel denotational test set.

Model Details

• Model type: RoBERTa-large sequence classifier
• Base model: roberta-large
• Task: Denotational relation classification between word-sense definitions
• Input: Two sense definitions for the same lexical item
• Output: A denotational relation label
• Language: English

Intended Use

This model is intended for research on lexical semantics, semantic change, polysemy, and sense-level semantic relations. It can be used to classify the relation between two definitions of a word sense, for example whether a newer meaning is a generalization, specialization, metaphorical extension, metonymic extension, homonym, or antonymic development of another meaning.

Example use cases include:

• studying semantic change in dictionaries or lexical resources;
• analyzing polysemous word senses;
• supporting sense-level semantic annotation;
• scaling exploratory studies of denotational change.

The model is not intended for high-stakes decision-making or for general-purpose natural language understanding outside lexical-semantic relation classification.

Input Format

The model expects a pair of definitions. In the experiments, definitions were concatenated and passed to the classifier.

Example:

from transformers import AutoTokenizer, AutoModelForSequenceClassification, pipeline

model_id = "ChangeIsKey/denotational-roberta-classifier"

classifier = pipeline(
    "text-classification",
    model=model_id,
    tokenizer=model_id
)

definition_1 = "the young of the domestic cow"
definition_2 = "the young of various large mammals"

text = definition_1 + " </s></s> " + definition_2
classifier(text)

Expected output:

[{"label": "generalization", "score": 0.XX}]

Training Data

The model was fine-tuned on a combined denotational-relation dataset referred to as WN+CN+UM, constructed from existing lexical-semantic resources:

• ChainNet for metaphor, metonymy, and homonymy relations;
• UniMet for additional metonymy examples;
• WordNet for generalization, specialization, and auto-antonymy examples.

These resources use definition or synset-pair information as proxies for relations between word senses.

Training Procedure

Definitions were concatenated and passed through roberta-large. A linear classification layer followed by a softmax activation was added on top to predict the relation label.

Training setup:

• maximum epochs: 10
• batch size: 8
• learning-rate schedule: linear
• warm-up: 10% of total training steps
• learning-rate search: {1e-4, 2e-4, 1e-5, 2e-5, 1e-6}
• model selection: best weighted F1 on the development set
• selected learning rate: 1e-6

Evaluation

The model was evaluated on:

1. the WN+CN+UM test set;
2. the SenseRel expert-annotated denotational dataset.

Dataset	Weighted F1
WN+CN+UM test set	0.734
SenseRel denotational dataset	0.684

On the SenseRel denotational dataset, this model achieved the best reported score among the evaluated systems.

Citation

If you use this model, please cite:

@inproceedings{cassotti-etal-2026-senserel,
    title = "{S}ense{R}el: A Sense-Level Benchmark for Denotational and Connotational Meaning Relations",
    author = "Cassotti, Pierluigi and
      Baes, Naomi and
      De Pascale, Stefano and
      de S{\'a}, J{\'a}der Martins Camboim and
      Periti, Francesco and
      Haslam, Nick and
      Geeraerts, Dirk and
      Tahmasebi, Nina",
    booktitle = "Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)",
    month = jul,
    year = "2026",
    address = "San Diego, California, United States",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2026.acl-long.20/",
    pages = "499--515"
}