Abstract: Importing test paper questions into the database is a key part of initializing the question bank. This thesis proposes a method based on SVM (Support Vector Machine) and shallow parsing to recognize massive test paper and automatically finish the initialization of the question bank. This approach first use SVM to build a hyperplane to separate test paper into two parts, which are the question numbers and the questions. Secondly, automata model based on the principle of shallow parsing is constructed to judge the question numbers which are recognized by SVM and revises the wrong results. Finally, the successful recognized questions are imported into the database automatically after confirming. A large number of experimental results demonstrate that this method does not need any artificial pre-processing work. It can be used to recognize the Word test paper that contains pictures, tables and formulas. The algorithm is proved to be feasible, effective and adaptable and the recognition rate can achieve 100%.