Data-Dependent Learning of Symmetric/Antisymmetric Relations for Knowledge Base Completion
Authors: Hitoshi Manabe, Katsuhiko Hayashi, Masashi Shimbo
AAAI 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our empirical evaluation shows that the proposed method outperforms the original Complex Embeddings and other baseline methods on the FB15k dataset. Experiments with synthetic data show that our method works as expected: Compared with the standard L1 regularization, the learned functions is more symmetric for symmetric relations and more antisymmetric for antisymmetric relations. Moreover, in KBC tasks on real datasets, our method outperforms the original Compl Ex with standard L1 and L2 regularization, as well as other baseline methods. |
| Researcher Affiliation | Collaboration | Hitoshi Manabe Nara Institute of Science and Technology Ikoma, Nara 630 0192, Japan manabe.hitoshi.me0@is.naist.jp Katsuhiko Hayashi NTT Communication Laboratories Seika-cho, Kyoto 619 0237, Japan hayashi.katsuhiko@lab.ntt.co.jp Masashi Shimbo Nara Institute of Science and Technology Ikoma, Nara 630 0192, Japan shimbo@is.naist.jp |
| Pseudocode | No | The paper describes mathematical update formulas for the training procedure but does not present them as structured pseudocode or an algorithm block. |
| Open Source Code | No | The paper does not contain an explicit statement or link indicating the availability of the source code for the described methodology. |
| Open Datasets | Yes | Following previous work, we used the Word Net (WN18) and Freebase (FB15k) datasets to verify the benefits of our proposed method. The dataset statistics are shown in Table 2. |
| Dataset Splits | Yes | The dataset statistics are shown in Table 2. Dataset |E| |R| #train #valid #test FB15k 14,951 1,345 483,142 50,000 59,071 WN18 40,943 18 141,442 5,000 5,000 |
| Hardware Specification | No | The paper does not provide specific details about the hardware used for running the experiments (e.g., CPU/GPU models, memory, or cloud instance types). |
| Software Dependencies | No | The paper describes algorithms used (e.g., RDA, Ada Grad) but does not list specific software dependencies with version numbers (e.g., programming languages, libraries, or frameworks). |
| Experiment Setup | Yes | For the multiplicative L1 regularizer, hyperparameters were set as follows: α = 1.0, λ = 0.05, η = 0.1. We selected the hyperparameters λ, α, and η via grid search such that they maximize the filtered MRR on the validation set. The ranges for the grid search were as follows: λ {0.01, 0.001, 0.0001, 0}, α {0, 0.3, 0.5, 0.7, 1.0}, η {0.1, 0.05}. During the training, learning rate η was tuned with Ada Grad (Duchi, Hazan, and Singer 2011), both for entity and relation vectors. The maximum number of training epochs was set to 500 and the dimension of the vector space was d = 200. The number of negative triplets generated per positive training triplet was 10 for FB15k and 5 for WN18. |