Learning Entity and Relation Embeddings for Knowledge Graph Completion

Authors: Yankai Lin, Zhiyuan Liu, Maosong Sun, Yang Liu, Xuan Zhu

AAAI 2015 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In experiments, we evaluate our models on three tasks including link prediction, triple classification and relational fact extraction. Experimental results show significant and consistent improvements compared to stateof-the-art baselines including Trans E and Trans H.
Researcher Affiliation Collaboration 1 Department of Computer Science and Technology, State Key Lab on Intelligent Technology and Systems, National Lab for Information Science and Technology, Tsinghua University, Beijing, China 2 Jiangsu Collaborative Innovation Center for Language Competence, Jiangsu, China 3 Samsung R&D Institute of China, Beijing, China
Pseudocode No The paper does not contain any pseudocode or algorithm blocks.
Open Source Code Yes The source code of this paper can be obtained from https: //github.com/mrlyk423/relation extraction.
Open Datasets Yes In this paper, we evaluate our methods with two typical knowledge graphs, built with Word Net (Miller 1995) and Freebase (Bollacker et al. 2008). Word Net provides semantic knowledge of words. In Word Net, each entity is a synset consisting of several words, corresponding to a distinct word sense. Relationships are defined between synsets indicating their lexical relations, such as hypernym, hyponym, meronym and holonym. In this paper, we employ two data sets from Word Net, i.e., WN18 used in (Bordes et al. 2014) and WN11 used in (Socher et al. 2013). WN18 contains 18 relation types and WN11 contains 11. Freebase provides general facts of the world. For example, the triple (Steve Jobs, founded, Apple Inc.) builds a relation of founded between the name entity Steve Jobs and the organization entity Apple Inc. In this paper, we employ two data sets from Freebase, i.e., FB15K used in (Bordes et al. 2014) and FB13 used in (Socher et al. 2013). We list statistics of these data sets in Table 1.
Dataset Splits Yes Dataset #Rel #Ent #Train #Valid # Test WN18 18 40,943 141,442 5,000 5,000 FB15K 1,345 14,951 483,142 50,000 59,071 WN11 11 38,696 112,581 2,609 10,544 FB13 13 75,043 316,232 5,908 23,733 FB40K 1,336 39528 370,648 67,946 96,678
Hardware Specification No The paper states:
Software Dependencies No The paper does not provide specific version numbers for software dependencies. It only mentions implementing models and using existing code like
Experiment Setup Yes For experiments of Trans R and CTrans R, we select learning rate λ for SGD among {0.1, 0.01, 0.001}, the margin γ among {1, 2, 4} , the dimensions of entity embedding k and relation embedding d among {20, 50, 100} , the batch size B among {20, 120, 480, 1440, 4800}, and α for CTrans R among {0.1, 0.01, 0.001}. The best configuration is determined according to the mean rank in validation set. The optimal configurations are λ = 0.001, γ = 4, k = 50, d = 50, B = 1440, α = 0.001 and taking L1 as dissimilarity on WN18; λ = 0.001, γ = 1, k = 50, d = 50, B = 4800, α = 0.01 and taking L1 as dissimilarity on FB15K. For both datasets, we traverse all the training triplets for 500 rounds.