Medical　ontology　learning　or　improving　is　automatically　learning　the　knowledge　in　ontology　format　from　medical　data,　mainly　text　data.　With　the　rise　of　the　word　vector　space,　improving　ontology　using　word　embedding　has　become　a　hot　spot.　Most　of　previous　studies　have　focused　on　how　to　acquire　different　ontological　elements　using　all　kinds　of　learning　technologies.　Few　studies　focus　on　the　prior　knowledge　in　a　given　ontology.　In　essence,　ontology　learning　or　improving　is　still　a　learning　process　based　on　existing　samples.　So,　the　type　and　number　of　knowledge　acquired　is　limited　by　existing　samples　in　a　given　ontology.　This　paper　firstly　formalizes　several　kinds　of　prior　knowledge　for　classes　in　a　given　ontology.　Then　we　propose　a　method,　named　improving　medical　ontology　based　on　word　embeddings　(IMO-WE),　to　enrich　different　types　of　knowledge　from　medical　text　according　to　characteristics　of　different　types　of　prior　knowledge.　At　last,　the　paper　collects　the　PubMed　Central　(PMC)　data　and　the　PHARE　ontology,　and　finishes　a　series　of　experiments　to　evaluate　the　IMO-WE.　The　experimental　results　yield　the　following　conclusions.　The　first　one　is　that　the　data-rich　model　can　achieve　higher　accuracy　for　the　IMO-WE　under　same　setting　in　training　progress.　So,　collecting　and　training　big　medical　data　is　a　viable　way　to　learn　more　useful　knowledge.　The　second　one　is　that　the　IMO-WE　can　be　used　to　improving　ontology　knowledge　when　medical　data　is　sufficiently　abundant　and　the　ontology　has　appropriate　prior　knowledge.　Moreover,　in　the　task　of　improving　synonymous　labels　through　similarity　distance,　the　accuracy　of　IMO-WE　is　significantly　better　than　that　of　the　Random　indexing　method.