Keyword　extraction　is　a　basic　text　retrieval　technique　in　natural　language　processing,　which　can　highly　summarize　text　content　and　reflect　the　author＇s　writing　purposes.　It　plays　an　important　role　in　document　retrieval,　text　classification　and　data　mining.　In　this　paper,　we　propose　a　TextRank　algorithm　based　on　PMI　(pointwise　mutual　information)　weighting　for　extracting　keywords　from　documents.　The　initial　transition　probability　of　the　candidate　words　is　constructed　by　calculating　the　PMI　between　vocabularies,　which　is　used　for　iterative　calculation　of　the　vocabulary　graph　model　within　TextRank　and　keyword　extraction.　Taking　into　account　the　mutual　information　between　the　vocabulary　in　the　document　set,　the　word　relationship　in　the　single　document　is　corrected,　which　is　helpful　to　improve　the　accuracy　of　document　keyword　extraction.　Experiments　show　that　our　method　achieves　better　performance　in　extracting　keywords　in　large-scale　text　data.　©　2019　IEEE.