Coreference Resolution in spaCy

2022-11-02 · Source: Explosion · Developer tools and consulting for AI, Machine Learning and NLP - Explosion.ai · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering · Depth: Intermediate, long

Summary

spaCy has released an experimental "coref" component (version 0.6 of "spacy-experimental") for coreference resolution, enabling machines to identify and link different expressions referring to the same entity across sentences. This is crucial for applications like dialogue systems. A pre-trained English "coref" model, "coreference_web_trf", is available. The full "coref" pipeline comprises five components, trained in two stages: a "coref resolver" pipeline (transformer + coref component) and a "span resolver" pipeline (transformer + sentencizer + span resolver component), then assembled with a "span cleaner". The article details setting up the project, preparing OntoNotes data, training both pipelines (the "coref resolver" took about two hours on a GPU, tracking six metrics including f-score), assembling the full model, and evaluating it using metrics like LEA. The trained model stores coreference information in Doc.spans, facilitating practical reference resolution in text.

Key takeaway

For NLP Engineers building dialogue systems or advanced text understanding applications, spaCy's experimental "coref" component provides a robust solution for resolving entity references. You should consider integrating "spacy-experimental==0.6" and the "coreference_web_trf" model to enhance contextual understanding in your systems. Explore the project's training methodology with OntoNotes as a template, or adapt it for your custom coreference annotations, to improve model accuracy and user experience.

Key insights

spaCy's experimental "coref" component enables robust coreference resolution for NLP applications.

Principles

• Coreference resolution is vital for NLP systems to understand context.
• Complex NLP pipelines can be modularized into sub-pipelines for training.
• GPU acceleration significantly reduces training time for transformer-based models.

Method

Train a "coref resolver" (transformer + coref) and a "span resolver" (transformer + sentencizer + span resolver) separately. Assemble them with a "span cleaner" into a full pipeline.

In practice

• Install "spacy-experimental==0.6" to access the "coref" component.
• Use "spacy project clone experimental/coref" to get the project template.
• Load "coreference_web_trf" for pre-trained English coreference.

Topics

• Coreference Resolution
• spaCy
• Natural Language Processing
• Transformer Models
• OntoNotes Dataset
• NLP Pipelines

Best for: NLP Engineer, Machine Learning Engineer, AI Engineer

Related on AIssential

• End-to-end Neural Coreference Resolution in spaCy — spaCy now features an experimental end-to-end neural coreference resolution component, enhancing natural language understanding.
• Coreference annotation for the characterization of characters in Portuguese literary works — Coreference annotation enhances character characterization in Portuguese literary corpora for computational literary studies.
• Universal Dependencies v2.5 Benchmarks for spaCy — spaCy v3.2 demonstrates competitive performance against Stanza and Trankit on Universal Dependencies v2.5 benchmarks using CoNLL 2018 evaluation.
• STPar: A Structure-Aware Triaffine Parser for Screenplay Character Coreference Resolution — STPar enhances movie screenplay coreference resolution by integrating structural, triaffine, and multi-task learning approaches.
• 3 SpaCy Tricks for Efficient Text Processing & Entity Recognition — Optimizing spaCy pipelines requires tailoring component usage, utilizing batch processing, and integrating rule-based entity recognition.
• CorPipe at CRAC 2026: Empty Nodes and Cross-Lingual Transfer in Multilingual Coreference Resolution — CorPipe 26 integrates empty node prediction with mention and coreference resolution for improved multilingual performance.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Explosion · Developer tools and consulting for AI, Machine Learning and NLP - Explosion.ai.