Spacy-GlossBert

A spaCy pipeline component for word sense disambiguation using the GlossBERT model.

Overview

This package provides a spaCy component that performs Word Sense Disambiguation (WSD) using the GlossBERT model. GlossBERT leverages BERT's contextual embeddings to disambiguate word senses by comparing the context with WordNet sense definitions (glosses).

Installation

pip install spacy-glossbert

You'll also need to download the spaCy model:

python -m spacy download en_core_web_sm

Usage

Basic Usage

import spacy
from spacy_glossbert import has_glossbert_wsd, get_synset_info

# Load spaCy with the GlossBERT component
nlp = spacy.load("en_core_web_sm")
nlp.add_pipe("glossbert_wsd", last=True)

# Process a text
doc = nlp("He went to the bank to deposit money.")

# Check if the document has been processed with GlossBERT
if has_glossbert_wsd(doc):
    # Print disambiguated senses
    for token in doc:
        synset = token._.glossbert_synset
        if synset:
            print(f"{token.text}: {token.pos_} -- {synset.name()} - {synset.definition()}")

    # Alternative: Get all sense information as a list of dictionaries
    senses = get_synset_info(doc)
    for sense in senses:
        print(f"{sense['text']}: {sense['synset']} - {sense['definition']}")

Visualization

The package includes visualization utilities using spaCy's displaCy:

from spacy_glossbert import visualize_wsd

# Visualize the disambiguated senses
visualize_wsd(doc)

Configuration Options

# Configure the component
nlp.add_pipe(
    "glossbert_wsd",
    config={
        "pos_filter": ["NOUN", "VERB", "ADJ"],  # Part-of-speech tags to process
        "supervision": True,  # Highlight the target word in context
        "model_name": "kanishka/GlossBERT"  # HuggingFace model name/path
    }
)

How It Works

The GlossBERT component:

1. Identifies tokens with POS tags specified in pos_filter
2. Retrieves WordNet synsets (possible senses) for each token
3. For each candidate sense, creates an input of the form: "context [SEP] definition"
4. Scores each sense using the GlossBERT model
5. Assigns the highest-scoring sense to each token

License

This project is licensed under the GNU General Public License v2.0 (GPLv2).

Credits

• GlossBERT model: Huang et al
• spaCy: Explosion AI
• WordNet: Princeton University

Author

Igor Morgado morgado.igor@gmail.com