There are many future directions in the knowledge graph space:

  1. Complex Reasoning

  2. Unified Framework

  3. Interpretability

  4. Scalability

  5. Knowledge Aggregation

  6. Automatic Construction and Dynamics

Complex Reasoning

Embedding methods tend to have limitation on complex logical reasoning and so relational path and symbolic logic are currently being heavily explored. Some upcoming research for handling complex reasoning are:

  1. Recurrent relational path encoding

  2. GNN-based message passing over knowledge graph

  3. Reinforcement learning-based pathfinding and reasoning

Unified Framework

Most work has been focusing on tackling knowledge acquisition knowledge graph completion and relation extraction as two separate tasks. Recent research work is moving in the direction of joint learning framework with information sharing between knowledge graph and text. A unified understanding of knowledge representation and reasoning is less explored.

Interpretability

For most real-world applications, interpretability of knowledge representation is extremely important. Some work focuses on combining neural networks with symbolic reasoning using logical rules to increase interpretability. With high interpretability, predictions have more credibility and transparent, making people more likely to trust and use it.

Scalability

There is a constant trade-off between computational efficiency and model expressiveness. Most work focuses on knowledge graph below 1 million entities. Many methods have been proposed to increase the scalability but most still struggle to scale to millions of entities and relations.

Knowledge Aggregation

The aggregation and fusion of global knowledge is essential to any knowledge-aware applications. Multi-modal information such as structured, semi-structured, and unstructured should be integrated into knowledge graph.

Automatic Construction and Dynamics

Most knowledge graphs are constructed manually, which is labour intensive and expensive. In order for knowledge graph to be widely adopted, we need a way to automate the construction of knowledge graphs from unstructured content. Recent work has been moving in this direction slowly in terms of semi-automatic construction using existing knowledge graphs as supervision. However, due to multimodality, heterogeneity, and large-scale applications, this is still a very difficult problem.

Finally, a dynamic knowledge graph is where we want to move towards where facts evolve over time just as how human knowledge tends to be. This is possible by exploring temporal information.

Ryan

Ryan

Data Scientist

Leave a Reply