With　the　construction　of　the　electronic　medical　record　system,　medical　record　data　begins　to　accumulate,　and　how　to　extract　essential　information　from　these　resources　has　become　a　concern.　And　named　entity　recognition(NER)　is　the　first　step.　With　the　help　of　doctors,　we　built　a　small　Chinese　electronic　medical　record　annotation　corpus.　But　the　NER　supervision　method　requires　a　large　amount　of　manually　labeled　corpus.　So　to　reduce　the　cost　of　it　and　make　better　use　of　the　unlabeled　corpus,　this　paper　proposes　a　semi-supervised　Chinese　electronic　medical　record　NER　model　based　on　ALBERT-BiLSTM-CRF　which　named　CEMRNER.　The　model　uses　a　Bidirectional　Long　Short　Term　Memory　network　(BiLSTM)　and　a　Conditional　Random　Field　model　(CRF)　to　train　the　data　and　introduces　the　pre-training　language　model　ALBERT　to　solve　the　problem　of　Chinese　representation.　At　the　same　time,　we　propose　a　dual　selected　strategy　to　select　the　high　confidence　samples　and　expand　the　training　set.　The　dual　strategy　can　ensure　the　accuracy　i　automatically　labeled　data,　and　reduce　the　error　iteration　in　semi-supervised　learning.　The　experiment　and　analysis　show　that　compared　with　other　models,　this　method　is　more　accurate　and　comprehensive.　The　precision,　recall　rate,　and　F1Score　are　85.45%,　87.81%,　and　86.61%,　respectively.　The　paper　proves　that　using　a　semi-supervised　method　and　pre-training　ALBERT　can　improve　the　accuracy　of　recognition　under　the　condition　of　less　labeled　data.　©　2020　ACM.