Open Relation and Event Type Discovery with Type Abstraction
This addresses the limitation of closed-world information extraction systems that rely on predefined ontologies, enabling adaptation to new domains, though it is incremental as it builds on existing type discovery methods.
The paper tackles the problem of automatically inferring new relation and event types from text corpora, which conventional closed-world information extraction approaches fail to do in new domains, by introducing type abstraction and a co-training framework, resulting in consistent advantages across multiple datasets.
Conventional closed-world information extraction (IE) approaches rely on human ontologies to define the scope for extraction. As a result, such approaches fall short when applied to new domains. This calls for systems that can automatically infer new types from given corpora, a task which we refer to as type discovery. To tackle this problem, we introduce the idea of type abstraction, where the model is prompted to generalize and name the type. Then we use the similarity between inferred names to induce clusters. Observing that this abstraction-based representation is often complementary to the entity/trigger token representation, we set up these two representations as two views and design our model as a co-training framework. Our experiments on multiple relation extraction and event extraction datasets consistently show the advantage of our type abstraction approach. Code available at https://github.com/raspberryice/type-discovery-abs.