Existing　named　entity　recognition　methods　are　often　based　on　large　training　samples　and　cannot　effectively　recognize　fine-grained　domain　entities　with　small　sample　sizes　In　order　to　solve　this　problem,　this　paper　proposes　an　unsupervised　method　based　on　contextual　domain　relevance　for　recognizing　biomedical　named　entities　with　small　sample　sizes.　Based　on　the　distributed　semantic　model,　the　statistical　and　linguistic　features　of　candidate　entities　in　corpora　are　described　by　using　occurrence　frequencies　of　contexts　of　candidate　entities.　Furthermore,　the　entity-corpus　relevance　assumption,　the　log-likelihood　ratio　and　the　domain-dependent　function　are　adopted　for　recognizing　objective　entities.　Experimental　results　show　that,　the　proposed　method　can　effectively　reduce　manual　interventions　and　improve　the　precision　rate　and　recall　rate　of　small-sample　biomedical　named　entity　recognition.　©　2019　IEEE.