Translating Embeddings for Knowledge Graph Completion with Relation Attention Mechanism
Authors: Wei Qian, Cong Fu, Yu Zhu, Deng Cai, Xiaofei He
IJCAI 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments show that our approach outperforms stateof-the-art methods significantly on public datasets, and our method can learn the true attention varying among relations. We evaluate our models with the tasks of link prediction and triple classification on public benchmark datasets: WN18, FB15k, WN11, and FB13. Experimental results show that our approach outperforms state-of-the-art models significantly. |
| Researcher Affiliation | Collaboration | State Key Lab of CAD&CG, College of Computer Science, Zhejiang University, China Alibaba-Zhejiang University Joint Institute of Frontier Technologies Fabu Inc., Hangzhou, China {qwqjzju, fc731097343}@gmail.com, {zhuyu cad, dcai}@zju.edu.cn, xiaofeihe@fabu.ai |
| Pseudocode | No | The paper does not contain a clearly labeled pseudocode or algorithm block. The methodology is described using mathematical equations and textual explanations. |
| Open Source Code | No | The paper mentions "1We use the code publicly available at https://github.com/thunlp/KB2E" but this refers to the code for Trans R, a baseline method, not the authors' own implementation of Trans At. There is no explicit statement about releasing the source code for their proposed method. |
| Open Datasets | Yes | Our experiments are carried out on public knowledge graphs dataset retrieved from Word Net [Miller, 1995] and Freebase [Bollacker et al., 2008] corpus. Specifically, WN11 [Socher et al., 2013] and WN18 [Bordes et al., 2013] are knowledge graphs retrieved from Word Net, while FB15k [Bordes et al., 2013] and FB13 [Socher et al., 2013] are extracted from Free Base. All these datasets consist of training set, validation set and testing set which are well organized. Table 1 lists the data statistics of the four datasets. |
| Dataset Splits | Yes | All these datasets consist of training set, validation set and testing set which are well organized. Table 1 lists the data statistics of the four datasets (WN11, WN18, FB13, FB15k), including #Train, #Valid, and #Test counts for each. We choose the optimal configuration according to Mean Rank on valid set and they are as follows: γ = 6, lr rate = 0.01, k = 100, α = 1, c = 10 on WN18 and γ = 2, lr rate = 0.001, k = 200, α = 1, c = 200 on FB15k. |
| Hardware Specification | No | The paper states: "The experiments are supported by Chengwei Yao in the Experiment Center of the College of Computer Science and Technology, Zhejiang University." This acknowledges support but does not provide specific details about the hardware (e.g., GPU models, CPU types, memory) used for running the experiments. |
| Software Dependencies | No | The paper mentions using Kmeans [Pedregosa et al., 2011], Xavier initialization [He et al., 2015a], and SGD [Bottou, 2010], but it does not specify version numbers for any software libraries or frameworks (e.g., Python, PyTorch, TensorFlow, scikit-learn) used in their implementation. |
| Experiment Setup | Yes | We select the margin γ among {1, 2, 4, 6}, the learning rate lr rate for SGD among {0.01, 0.001}, the dimension of vectors k among {50, 100, 200}, the loss weight α among {0.1,1,10}. We just adjust mini-batch size to make each epoch has 100 mini-batch and run training process with 1000 epochs. We reset attention every 100 epochs. The cluster number of Kmeans c is set according to the number of relations. The optimal configuration according to Mean Rank on valid set and they are as follows: γ = 6, lr rate = 0.01, k = 100, α = 1, c = 10 on WN18 and γ = 2, lr rate = 0.001, k = 200, α = 1, c = 200 on FB15k. |