Detecting Hypernymy Relations Between Medical Compound Entities Based On Deep Learning

Named entities composed of multiple continuous words frequently occur in domain-specific knowledge graphs.In general,these named entities are composable and extensible,such as names of symptoms and diseases in the medical domain.Unlike the general entities,we address them as compound entities,and try to identify hypernymy relations between them.Hypernymy detection between compound entities plays a critical role in domain-specific knowledge graph construction.Traditional hypernymy detection approaches do not perform well on compound entities for two reasons.One is the lack of contextual information,and the other is the absence of compound entities,i.e.out-of-vocabulary(OOV)problem.In this paper,we propose a hybrid-attention-based method called HA-Bi-GRU-CapsNet for the detection of hypernymy relations.The hybrid attention mechanism consists of a heuristic attention and a self-adaptive attention,which are used for the lack of contextual information.The attentions focus on the differences of two compound entities on the lexical and semantic level,respectively.For OOV problem,the English words or Chinese characters in compound entities are fed into bidirectional gated recurrent units(Bi-GRUs).Additionally,we use capsule network(CapsNet)to determine the existence of hypernymy relations under different cases.We construct an English corpus and a Chinese corpus,respectively by extracting clinical finding pairs in SNOMED-CT and Chinese healthcare websites.We conduct experiments on them.Experimental results show that our proposed method outperforms other baseline methods on both English and Chinese corpora of symptom and disease pairs.We also experiment our attention mechanism on open datasets.Experimental results show that the effect of our attention mechanism is decided by the type of entities.So it is necessarily to identify the characteristics of entities,before applying the attention mechanism.