

# Problem Statement

Existing KG embedding models merely focus on representing of an ontology view for abstract and commonsense concepts or an instance view for special entities that are instantiated from ontological concepts.

# Challenge

• mappings difficult :the semantic mappings from entities to concepts and from relations to meta-relations are complicated and difficult to be precisely captured by any current embedding models
• the scales and topological structures are different in ontological views and instance views

# Introduction

• instance embeddings provide detailed and rich information for their corresponding ontological concepts.
• a concept embedding provides a high-level summary of its instances, which is extremely helpful when an instance is rarely observed.

# contribution

• a novel KG embedding model named JOIE, which jointly encodes both the ontology and instance views of a KB
• cross-view association model : a novel KG embedding model named JOIE, which jointly encodes both the ontology and instance views of a KB
• cross-view grouping technique : assumes that the two views can be forced into the same embedding space
• cross-view transformation technique : enables non-linear transformations from the instance embedding space to the ontology embedding space
• intra-view embedding model : characterizes the relational facts of ontology and instance views in two separate embedding spaces
• three state-of-the-art translational or similarity-based relational embedding techniques
• hierarchy-aware embedding: based on intra-view non- linear transformations to preserve ontologies hierarchical substructures.
• implement two experiments:
• the triple completion task : confirm the effectiveness of JOIE for populating knowledge in both ontology and instance-view KGs, which has significantly outperformed various baseline models.
• the entity typing task : show that JOIE is competent in discovering cross-view links to align the ontology-view and the instance-view KGs.

## Intra-view Model

intra损失函数：

### Hierarchy-Aware Intra-view Model for the Ontology

• $J_{\text { Intra }}^{\mathcal{G} o} \backslash \mathcal{T}$: 默认的视图内模型的丢失，该模型仅在具有规则语义关系的三元组上训练
• $J_{\text { Intra }}^{\mathrm{HA}}$明确训练三元组与形成本体层次结构的元关系

# EXPERIMENTS

## Case Study

### Long-tail entity typing

In KGs, the frequency of entities and relations often follow a long-tail distribution (Zipf’s law)

# FUTURE WORK

• Particularly, instead of optimizing structure loss with triples (first-order neighborhood) locally, we plan to adopt more complex embedding models which leverage information from higher order neighborhood, logic paths or even global knowledge graph structures.
• We also plan to explore the alignment on relations and meta-relations like entity-concept.
• exploring different triple encoding techniques
• Note that we are also aware of the fact that there are more comprehensive properties of relations and meta-relations in the two views such as logical rules of relations and entity types. Incorporating such properties into the learning process is left as future work.

# 思考

-------------本文结束感谢您的阅读-------------