Open knowledge enrichment for long-tail entities

Proposed OKELE, an end-to-end method to enrich long-tail entities from the open Web! It’s based on the idea that we can infer missing knowledge of long-tail entities by looking at similar popular entities. OKELE enrich long-tail entities by searching for a set of true facts about it on the Web and it does so as follows:

  1. Constructed a novel entity property prediction model based on GNN and attention mechanism to predict the missing properties of long-tail entities. The model will then look for values for these missing properties

  2. Explored different semi-structured, unstructured, and structured Web sources and designed different extraction methods for it

  3. Present a new fact verification model based on probabilistic graphical model to resolve conflicting values / facts about the same properties from different sources

Future Work

  1. Optimise key modules to accelerate enrichment speed

  2. Explore a full neural network-based architecture

Adapting meta knowledge graph information for multi-hop reasoning over few-shot relations

Proposed Meta-KGR, which uses meta-learning to effectively learn meta parameters from high-frequency relations to improve predictions on low-frequency relations. In addition, the solution also showcase multi-hop reasoning, giving the reasoning path that the knowledge graph took to derive the final answer which improves interpretability. Previous work has deal with zero-shot or one-shot relations but there are two main downfalls:

  1. They lack interpretability as most of them uses embedding-based methods

  2. Focus too much on zero-shot or one-shot relations which are a little too far away from practical scenarios!

Overall, the Meta-KGR uses a reinforcement learning agent to identify target entities and reasoning paths, therefore improving both explainability and few-shot relations!

Bootstrapping entity alignment with knowledge graph embedding

Proposed a bootstrapping approach to create more training data for entity alignment. This involves iteratively labels entity alignment as training data and training a classifier to learn alignment-oriented KG embeddings. It also uses alignment editing method to reduce error accumulation during iterations. The two challenges with existing embedding-based entity alignment are:

  1. Alignment-oriented KG embeddings are largely unexplored despite many existing work focusing on it

  2. Embedding-based entity alignment relies on large amount of existing entity alignment training data. However, this is not always the case in the practical world and so embedding-based models tend to have low precision

Future Work

  1. Explore cross-lingual word embeddings for attribute values

  2. Use RNNs to model complex semantics of KGs

Ryan

Ryan

Data Scientist

Leave a Reply