Meta relational learning for few-shot link prediction in knowledge graphs

Proposed MetaR, Meta Relational Learning (MetaR) for few-shot link prediction in KGs. Our intuition is that the most important information to be transferred to do few-shot linking should be the common and shared knowledge within one task. This is called relation-specific meta information and it’s helpful in two ways:

  1. Transferring common relation information from observed to incomplete triples

  2. Accelerating the learning process within one task by only observing a few instances

Therefore, there are two types of relation-specific meta information:

  1. Relation meta. The high order information connecting head and tail entities

  2. Gradient meta. The loss gradient of relation meta which will be used to make update before transferring relation meta to incomplete triples during prediction Our approach is independent of the background knowledge graphs

Future Work

  1. Consider utilising other more valuable information about sparse entities for few-shot link predictions

Robust distant supervision relation extraction via deep reinforcement learning

Proposed a deep reinforcement learning framework to create a false-positive classifier to automatically identify false positives for each relation type, making distant supervision more robust. Those identified false positives are redistributed into the negative samples. Our proposed framework is model-independent, which means that it can be applied to any SOTA relation extractors.



Data Scientist

Leave a Reply