As　the　cases　exploded,　leading　legal　judgment　prediction　becomes　a　promising　application　of　artificial　intelligence　techniques　in　the　legal　field.　The　goal　of　legal　judgment　prediction　is　to　predict　the　judgment　results　based　on　the　facts　information　of　a　case.　However,　the　classifier　of　the　traditional　method　has　poor　accuracy　performance　and　cost　large　computational　time.　The　commonly　used　deep　learning　models　are　CNN　and　RNN.　In　this　paper,　CNN-BiGRU　was　established　and　analyzed,　which　combined　the　good　extraction　ability　of　CNN　for　local　feature　information　and　RNN　for　long-term　dependencies　information　of　the　text.　Compared　with　the　CAIL　2018　dataset,　the　prediction　accuracy　of　the　charges,　law　articles　and　the　terms　of　penalty　are　94.8%,　93.6%,　and　73.4%,　respectively.　Results　showed　that　CNN-BiGRU　has　a　higher　prediction　accuracy　than　CNN　or　RNN　alone　and　a　good　training　efficiency　over　baselines.　The　effectiveness　and　practicability　of　this　model　are　validated.　©　2020　ACM.