Validation of Growing Knowledge Graphs by Abductive Text Evidences
Authors: Jianfeng Du, Jeff Z. Pan, Sylvia Wang, Kunxun Qi, Yuming Shen, Yu Deng2784-2791
AAAI 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | To evaluate the proposed validation mechanism, four knowledge graphs with logical theories are constructed from the four great classical masterpieces of Chinese literature. Experimental results on these datasets demonstrate the efficiency and effectiveness of the proposed mechanism. |
| Researcher Affiliation | Collaboration | Guangdong University of Foreign Studies, Guangzhou 510420, P.R.China Department of Computing Science, The University of Aberdeen, Aberdeen AB24 3UE, UK IBM T.J. Watson Research Center, Yorktown Heights, NY 10598, USA |
| Pseudocode | Yes | Figure 1: Computing all TEP-based abductive text evidences |
| Open Source Code | No | The paper states 'All datasets mentioned in Table 1 and our constructed OWL 2 RL ontologies are available at http://dataminingcenter.net/papers/AAA1-19-data.zip.', but there is no explicit statement or link indicating that the source code for the methodology described in the paper is publicly available. |
| Open Datasets | Yes | All datasets mentioned in Table 1 and our constructed OWL 2 RL ontologies are available at http://dataminingcenter.net/papers/AAA1-19-data.zip. |
| Dataset Splits | No | We divided each knowledge graph into a training set and a test set, where triples in the test set are all treated as new triples. |
| Hardware Specification | Yes | The computation of TEP-based abductive text evidences was conducted in a laptop with 8GB memory and 2-core 2.5GHz CPU. The ranking of TEP-based abductive text evidences by our proposed method (CNNRP/PCNNRP) or a variant (CNNP/PCNNP) needs to learn a neural model in Figure 2 from the TE-training set and then apply the learnt model to compute certain values used in Formula (8) or Formula (9). These experiments were conducted in a workstation with 64GB memory and 28-core 2.2GHz CPU. |
| Software Dependencies | No | We implemented the proposed methods for computing and ranking TEP-based abductive text evidences in Java... Both ranking methods employ Adam (Kingma and Ba 2014) as the stochastic optimization algorithm in the training course... Every logical theory is originally expressed in OWL 2 RL (Grau et al. 2008), a tractable profile of OWL 2 for modeling ontologies, and then translated to a Horn theory by standard transformation. Every constructed OWL 2 RL ontology is rather complex. It contains transitivity axioms, such as one axiom declaring that relatives are transitive, as well as property chain axioms, such as another axiom declaring that daughters in law are wives of sons. We did not split a sentence window into words because existing word segmentation tools do not work well for ancient Chinese sentences in the masterpieces. |
| Experiment Setup | Yes | For training the model we uniformly set the dimension for word embeddings as 100, the dimension for position embeddings as 10, the dimension for sentence embeddings as 100 for CNN and 150 for PCNN, the window size in CNN/PCNN as 2, the initial learning rate as 0.001 for Adam (Kingma and Ba 2014), the probability for applying dropout (Srivastava et al. 2014) as 0.1, and the learning epochs as 20. |