Abstract
Taxonomy expansion is critical for maintaining up-to-date knowledge graphs, yet most existing methods treat node addition ("Attach") and synonym identification ("Merge") as separate tasks. We present TEAM (Taxonomy Expansion with Attach and Merge), the first multi-task learning framework designed to handle both operations simultaneously. By utilizing neighborhood-aware "ego-trees" and position-enhanced Graph Attention Networks (GAT), TEAM captures local hierarchical context more effectively than previous approaches. Our experiments on low-resource IndoWordNet taxonomies (Assamese, Bengali, and Hindi) demonstrate that TEAM significantly outperforms state-of-the-art baselines, particularly in ranking accuracy and robustness to out-of-vocabulary terms.
Introduction
Taxonomies like WordNet provide essential structured relationships for NLP tasks but often suffer from limited coverage due to the high cost of manual curation. Automatic taxonomy expansion aims to integrate new concepts into an existing hierarchy. This process involves two distinct operations: Attach, where a new leaf node is added under an appropriate parent, and Merge, where a new term is identified as a synonym for an existing node.
While previous research has addressed these tasks in isolation, TEAM treats them as a unified problem. This approach leverages shared semantic information, as the features required to identify a synonym often overlap with those needed to find a correct parent.
Methodology
The TEAM framework introduces two key innovations to represent nodes within the taxonomy:
- Ego-tree Representations: Instead of analyzing nodes in isolation, TEAM extracts an "ego-tree" for each candidate anchor node, consisting of the node itself along with its immediate parents and children. This provides the model with vital local hierarchical context.
- Position-Enhanced Graph Attention Networks (GAT): A GAT encodes these ego-trees, incorporating position embeddings to distinguish the roles of different nodes (e.g., parent vs. child) within the local structure.
The authors propose two architectural variants:
- TEAM-RG (Regression): Formulates expansion as a ranking task using InfoNCE loss. This contrastive learning approach forces the model to distinguish the correct anchor from thousands of negative candidates.
- TEAM-CL (Classification): Treats the problem as a three-way classification (Merge, Attach, or No-operation) combined with pairwise hinge loss for ranking.
Experimental Setup and Results
The model was evaluated on three low-resource taxonomies from the IndoWordNet family: Assamese, Bengali, and Hindi. These datasets present a significant challenge due to their smaller size compared to English WordNet.
TEAM-RG demonstrated substantial improvements over the previous state-of-the-art, Taxo-Expan. In the Assamese WordNet, TEAM-RG achieved a Mean Rank (MR) of 144.92, a significant leap from Taxo-Expan’s 341.81. For the merge operation, TEAM-RG achieved near-perfect Mean Reciprocal Rank (MRR) scores (~1.00) when definitions were provided. Furthermore, in experiments with out-of-vocabulary (OOV) Assamese words, TEAM-RG maintained its lead with a Mean Rank of 65.30 compared to 386.30 for the baseline.
Conclusion
TEAM demonstrates that multi-task learning for "attach" and "merge" operations provides a powerful synergy for taxonomy expansion. By focusing on structural context through ego-trees and utilizing contrastive learning, the model achieves superior performance even in low-resource settings. Future work includes addressing class imbalance in the classification variant and exploring advanced contextual encoders like BERT to further enhance cross-lingual performance.
Summary updated