Factorizing YAGO: scalable machine learning for linked data

Proposed tensor factorisation method to relational learning on Linked Open Data (LOD). Ontological knowledge can be incorporated into the factorisation to improve learning results. The methodology was able to factorise YAGO 2 ontology and globally predict statements for large knowledge base using a single dual-core desktop computer. This was possible due to a novel approach of taking advantage of the sparsity of LOD data when performing factorisation. The pare also extended RESCAL.

Future Work

  1. Reduce the effective sparsity of a tensor representation by adding typed relations to factorisation

  2. Using efficient methods to find good parameters value as doing cross-validation on large-scale data is costly! There are scalable Bayesian methods that could potentially help with this

Reasoning with neural tensor networks for knowledge base completion

Introduced Neural Tensor Network (NTN) for reasoning over relationships between two entities. The contributions are as follows:

  1. Proposed NTN for modelling relational data and also it generalises few previous neural network models

  2. Introduced a new way to represent entities, where we represent each entity as the average of its word vectors, capturing the sharing of words between entities

  3. Showcase that models improve when word vectors are initialised with vectors learned from large unsupervised corpus

Future Work

  1. Scaling the number of slices based on available training data for each relation

  2. Extend the paper’s ideas to reason over free text

Ryan

Ryan

Data Scientist

Leave a Reply