Graded Word Sense Disambiguation (WSD) Model

Model Summary

This model is a fine-tuned version of RoBERTa-Large for Graded Word Sense Disambiguation (WSD). It is designed to predict the degree of applicability (1-4) of a word sense in context by leveraging large-scale sense-annotated corpora. The model is based on the work outlined in:

Reference Paper: Pierluigi Cassotti, Nina Tahmasebi (2025). Sense-specific Historical Word Usage Generation.

This model has been trained to handle graded WSD tasks, providing continuous-valued predictions instead of hard classification, making it useful for nuanced applications in lexicography, computational linguistics, and historical text analysis.

Model Details

• Base Model: roberta-large
• Task: Graded Word Sense Disambiguation (WSD)
• Fine-tuning Dataset: Oxford English Dictionary (OED) sense-annotated corpus
• Training Steps:
  • Tokenizer augmented with special tokens (<t>, </t>) for marking target words in context.
  • Dataset preprocessed with sense annotations and word offsets.
  • Sentences containing sense-annotated words were split into train (90%) and validation (10%) sets.
  • Objective: Predicting a continuous label representing the applicability of a sense.
  • Evaluation Metric: Root Mean Squared Error (RMSE).
• Batch Size: 32
• Learning Rate: 2e-5
• Epochs: 1
• Optimizer: AdamW with weight decay of 0.01
• Evaluation Strategy: Steps-based (every 10% of the dataset).

Training & Fine-Tuning

Fine-tuning was performed using the Hugging Face Trainer API with a custom dataset loader. The dataset was processed as follows:

1. Preprocessing

   • Example sentences were extracted from the OED and augmented with definitions.
   • The target word was highlighted with special tokens (<t>, </t>).
   • Each instance was labeled with a graded similarity score.

2. Tokenization & Encoding

   • Tokenized with AutoTokenizer.from_pretrained("roberta-large").
   • Definitions were concatenated using the </s></s> separator for cross-sentence representation.

3. Training Pipeline

   • Model fine-tuned on the regression task with a single linear output head.
   • Used Mean Squared Error (MSE) loss.
   • Evaluation on validation set using Root Mean Squared Error (RMSE).

Usage

Example Code

from transformers import AutoModelForSequenceClassification, AutoTokenizer
import torch

tokenizer = AutoTokenizer.from_pretrained("ChangeIsKey/graded-wsd")
model = AutoModelForSequenceClassification.from_pretrained("ChangeIsKey/graded-wsd")

sentence = "The <t>bank</t> of the river was eroding due to the storm."
target_word = "bank"
definition = "The land alongside a river or a stream."

tokenized_input = tokenizer(f"{sentence} </s></s> {definition}", truncation=True, padding=True, return_tensors="pt")
with torch.no_grad():
    output = model(**tokenized_input)
    score = output.logits.item()

print(f"Graded Sense Score: {score}")

Input Format

• Sentence: Contextual usage of the word.
• Target Word: The word to be disambiguated.
• Definition: The dictionary definition of the intended sense.

Output

• A continuous score (between 1 and 4) indicating the similarity of the given definition with respect to the word in its current context.

Citation

If you use this model, please cite the following paper:

@article{10.1162/tacl_a_00761,
    author = {Cassotti, Pierluigi and Tahmasebi, Nina},
    title = {Sense-specific Historical Word Usage Generation},
    journal = {Transactions of the Association for Computational Linguistics},
    volume = {13},
    pages = {690-708},
    year = {2025},
    month = {07},
    abstract = {Large-scale sense-annotated corpora are important for a range of tasks but are hard to come by. Dictionaries that record and describe the vocabulary of a language often offer a small set of real-world example sentences for each sense of a word. However, on their own, these sentences are too few to be used as diachronic sense-annotated corpora. We propose a targeted strategy for training and evaluating generative models producing historically and semantically accurate word usages given any word, sense definition, and year triple. Our results demonstrate that fine-tuned models can generate usages with the same properties as real-world example sentences from a reference dictionary. Thus the generated usages will be suitable for training and testing computational models where large-scale sense-annotated corpora are needed but currently unavailable.},
    issn = {2307-387X},
    doi = {10.1162/tacl_a_00761},
    url = {https://doi.org/10.1162/tacl\_a\_00761},
    eprint = {https://direct.mit.edu/tacl/article-pdf/doi/10.1162/tacl\_a\_00761/2535111/tacl\_a\_00761.pdf},
}