Taxonomy Completion via Triplet Matching Network

Authors: Jieyu Zhang, Xiangchen Song, Ying Zeng, Jiaze Chen, Jiaming Shen, Yuning Mao, Lei Li4662-4670

AAAI 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments on four real-world large-scale datasets show that TMN achieves the best performance on both taxonomy completion task and the previous taxonomy expansion task, outperforming existing methods.
Researcher Affiliation Collaboration Jieyu Zhang ,1 Paul G. Allen School of Computer Science & Engineering, University of Washington, WA, USA 2 Department of Computer Science, University of Illinois Urbana-Champaign, IL, USA 3 Byte Dance AI Lab
Pseudocode No The paper provides architectural diagrams and mathematical formulations but no explicit pseudocode or algorithm blocks.
Open Source Code Yes 1The code is released at https://github.com/Jieyu Z2/TMN
Open Datasets Yes We study the performance of TMN on four large-scale real-world taxonomies. Microsoft Academic Graph (MAG). We evaluate TMN on the public Field-of-Study (Fo S) Taxonomy in Microsoft Academic Graph (MAG) (Sinha et al. 2015). ... Word Net. Based on Word Net 3.0, we collect verbs and nouns along with the relations among them to form two datasets which we refer to as Word Net-Verb and Word Net-Noun, respectively.
Dataset Splits Yes For each dataset, we randomly sample 1,000 nodes for validation and another 1,000 for test.
Hardware Specification No The paper does not provide specific hardware details (e.g., CPU, GPU models, memory) used for the experiments.
Software Dependencies No The paper mentions 'Adam optimizer' and links to 'pytorch.org/docs/stable/optim.html' implying PyTorch, but does not provide specific version numbers for PyTorch or any other software dependencies.
Experiment Setup Yes Parameter Settings. For learning-based methods, we use Adam optimizer with initial learning rate 0.001 and Reduce LROn Plateau scheduler3 with ten patience epochs. During model training, the batch size and negative sample size is set to 128 and 31, respectively. We set k, i.e., the dimension of internal feature representation, to be 5. For TMN, we simply set λ1 = λ2 = λ3 = 1 to avoid heavy hyperparameter tuning.