EUSKOR: End-to-end coreference resolution system for Basque

Area of Science:

• Computational Linguistics
• Natural Language Processing
• Language Technology

Background:

• Coreference resolution is vital for understanding text.
• Adapting existing NLP tools to low-resource languages like Basque presents unique challenges.
• The Stanford Coreference module is a widely used system.

Purpose of the Study:

• To adapt the Stanford Coreference resolution module for the Basque language.
• To develop an end-to-end coreference resolution system for Basque.
• To analyze the impact of language-specific features and preprocessing on coreference resolution performance.

Main Methods:

• Adaptation of the Stanford Coreference module, considering Basque linguistic characteristics.
• Integration into a linguistic analysis pipeline for an end-to-end system.
• Experimentation comparing automatic mentions vs. gold mentions and analyzing sieve contributions.

Main Results:

• Achieved a 7.07 CoNLL score gain over the baseline system by incorporating language-specific features.
• Demonstrated a significant performance increase (11.5 CoNLL points) when using gold mentions compared to automatic mentions.
• Identified morphology as essential for agglutinative languages, leading to the addition of two new sieves that improved F1 scores.

Conclusions:

• Language-specific adaptations, particularly morphology, are critical for effective coreference resolution in agglutinative languages.
• The developed system provides a valuable resource for Basque NLP.
• The adaptation methodology can serve as a blueprint for other low-resource languages.