Ontology Knowledge-enhanced In-Context Learning for Action-Effect Prediction

Implementing human-level reasoning about action effects is an important competence for a cognitive agent: given precondition and action descriptions, a system should be able to infer the change in the physical world that the action causes. In this work, we propose a new action-effect prediction task. We explore few-shot learning with large pre-trained language models based on a limited number of samples and propose task-relevant ontology knowledge (from KnowRob ontology) integration for in-context learning with generative pre-trained transformer (GPT) models. Specifically, we develop an ontology-to-text transformation to bridge the gap between symbolic knowledge and text. We further introduce unseen knowledge learning via GPT-3 to infer knowledge for concepts that do not have definitions in the knowledge base. We evaluate our proposed method on two human-annotated datasets. Experimental results demonstrate that our approach can improve the performance of large-scale, state-of-the-art models on two action-effect prediction datasets.