|
Knowledge extraction is a significant issue in fields including knowledge graph construction and natural language processing. Triple extraction, as a crucial issue within knowledge extraction, is intended to acquire structured triples from unstructured texts, that is, to identify entities, their types and relations. Typically, it provides technical support for downstream tasks such as information recommendation, semantic search, and question answering. However, present triple extraction methods fail to sufficiently utilize the information related to relation types and entity types during the sentence embedding generation, and cannot fully exploit the auxiliary semantic information about relation for entity extraction. To address the aforementioned challenges, this paper proposes a triple extraction approach founded on Meta-type Prompt learning and Bidirectional relation Collaborative Attention (MPBCA). That method utilizes meta-type prompt learning which introduces relation types, entity types and their correlations to optimize the token embedding. Thereby, the generated word embeddings can be more adaptive for the triple extraction task by exploring the intrinsic logical connections of target triple components. Furthermore, our approach designs a bidirectional relation complementary attention mechanism to identify entity head and tail positions. That mechanism not only strengthens the semantic modelling capabilities for relations between entities, but also improves fault tolerance and captures the potentially accurate target triples by using bidirectional sentence structures, not unidirectional structures. The results of experiments on two public datasets show that our triple extraction approach MPBCA proposed in this paper is superior to the existing methods, confirming the effectiveness and superiority of the proposed model.
|
