A hierarchical framework for relation extraction with reinforcement learning

Presents a novel paradigm in dealing with relation extraction by treating related entities as arguments of a relation. We apply hierarchical reinforcement learning (HRL) framework to enhance the interaction between entity mentions and relation types for the relation extraction tasks. The whole extraction process has two-level RL policies; one for relation detection and the other for entity extraction. This means that we have:

  1. High-level reinforcement learning process to detect relation indicators

  2. Low-level RL process to detect participating entities for the detected relation

This makes the whole extraction process more like a sequential scan, where it will go through the each sentence multiple times! This approached solve two main issues:

  1. Previous work determine a relation type ONLY after all entities have been recognised and so the interactions between the two tasks entity and relation extraction) is not fully captured. This was solved using our solution of treating entities as the arguments of a relation, creating this dependency

  2. Overlapping relations (one-to-many) is still a problem in the relation extraction task where one entity may be involved in multiple relations in the same sentence. This was addressed by our hierarchical structure, allowing us to extract and handle multiple relations in a sentence separately and sequentially!

Future Work

  1. Generalise the hierarchical extraction framework to many other pairwise or triple-wise extraction tasks such as aspect-opinion mining or ontology induction

Few-shot knowledge graph completion

Proposed FSRL, few-shot relation learning model, that aims at discovering facts of new relations with few-shot references. FSRL can effectively capture knowledge from KGs, aggregate representations of few-shot references, and match similar entity pairs of reference set for every relation. The contributions are as follows:

  1. Introduced a new few-shot KG completion problem that’s more suitable for practical scenarios

  2. Proposed a few-shot learning relation learning model to solve the problem

  3. Achieved SOTA results with two public datasets

The FSRL learned model can infer true entity pairs for any new relation without additional fine-tuning steps! The model has two main steps:

  1. Proposed a relation-aware heterogeneous neighbour encoder to learn entity embeddings based on graph structure and attention mechanism

  2. Design a recurrent auto-encoder aggregation network to model the interactions of few-shot reference entity pairs and accumulate their expression capabilities for each relation

Future Work

  1. Utilise a better model training process such as model-agnostic meta-learning or incorporating contextual information such as entity attributes or text description to improve quality of entity embeddings

Ryan

Ryan

Data Scientist

Leave a Reply