Collective Multi-type Entity Alignment Between Knowledge Graphs

Proposed CG-MuAlign, a Collective-Graph neural network for Multi-type entity Alignment. In a knowledge graph, there are different entity types and relation types and given the nature of entity and relation types, it makes sense to have a different alignment strategy for the different entity types. Therefore, a multi-type entity alignment algorithm is needed for effective knowledge integration! Our GNN-based KG entity alignment framework can collectively align entities of different types and can make predictions on unseen entities using attention mechanism and neighbourhood information. Our framework can scale up to large-scale KGs by avoiding costly computation in each layer of the GNN and by relation-aware neighbourhood sampling! Previous entity alignment methods have two main problems:

  1. Transductive —> Inductive. Previous methods only focus on related entities and ignored the rich attribute information that entities hold. In addition, when new entities get added to the graph, the methods will need to be retrain which it’s costly

  2. Labelled type —> Unlabelled type. Previous methods tend to perform well for entity types with large training data and failed on entity types with sparse data. Intuitively, rich connections between entity types should help boost performance on entity types with small training data but connections are not being used effectively yet.

Future Work

  1. Extend CG-MuAlign to transductive setting, where we need to make collective decisions on structural information

  2. Handle multiple (> 2) knowledge graphs alignment simultaneously

  3. Come up with a unified GNN framework for jointly aligning entity and relation together

Relation extraction using supervision from topic knowledge of relation labels

Mine topic knowledge of a relation to explicitly represent the semantics of the relation and model relation extraction as a matching problem. The paper use topic knowledge to improve RE in three aspects:

  1. Deep sentence-relation matching. Model RE as a matching problem where given an entity pair, the input is a sentence-relation pair and the output is a matching score where the relation is represented by its topic words

  2. Sample reweighting. A new reweighting scheme based on topic knowledge to highlight the high-quality samples with a large weight in the training process to pre-estimate the semantic distance between the sentence and the topic knowledge of the relation using WMD

  3. Proposed two knowledge-guided negative sampling strategies based on the semantic distance using topic knowledge.

Future Work

  1. Apply our framework to other classification-based NLP tasks

Ryan

Ryan

Data Scientist

Leave a Reply