A Unified Taxonomy-Guided Instruction Tuning Framework for Entity Set Expansion and Taxonomy Expansion

Entity set expansion, taxonomy expansion, and seed-guided taxonomy construction are three representative tasks for automatically enriching an existing taxonomy with emerging concepts.Previous studies have treated them as separate tasks, leading to techniques that are specialized for one task but lack generalizability and a holistic perspective.In this paper, we propose a unified solution to address all three tasks.Specifically, we identify two fundamental skills facilitating the three tasks: finding "siblings" and finding "parents".To this end, we introduce a taxonomy-guided instruction tuning framework that trains a large language model to generate siblings and parents for query entities, where the joint pretraining process enables mutual reinforcement of these two skills.Extensive experiments on multiple benchmark datasets validate the effectiveness of our proposed TAXOINSTRUCT framework, demonstrating its superiority over task-specific baselines across all three tasks.Our codes and data are available at https: //github.com/yanzhen4/TaxoInstruct.Instruction: Given a list of entities, output the most likely parent class for the entity given by user.Input: Find the parent class for {Heart Enlargement, Heart Rupture, Heart Defect}.Instruction: Given a category and an entity set belonging to this category, output other entities belonging to this category and sharing the same granularity as the seeds.Input: Find other entities belonging to the category Heart Disease and sharing the same granularity as the seeds {Heart Enlargement, Heart Rupture, Heart Defect}. Entity Set ExpansionInstruction: Given a set of candidate parent classes and an entity, output the most likely parent class for the entity given by user.Input: Given candidate parents