| There is a software product which uses Oracle as its database. But during the sales process, a lot of more and more customer demand to use SQL-Server instead of Oracle. Because there are tremendous amount of SQL codes used the current product. The migration of the DB should transfer all these SQL codes to those which can be run on SQL-Server. The workload is so large that we have to consider doing these transfers by the help of the computer. After having some try on this idea, we find it’s reasonable to translate these SQL codes by using the natural language process for reference.By learning lessons from the natural language process, we implement a SQL translation tool. At first, the tool uses SRILM to count the words frequency and estimate the language model. Then, the tool can use the k-best algorithm to parse the sentence to the phrases. After that, the program can use the language model trained by the bilingual corpus to translate the phrases to T-SQL language by comparing the context similarity. The last step of the translation is searching the best sequence of the translated phrases by solving the Traveling salesman problem with Held&Karp DP algorithm. For those are not translated properly or ordered in a valid sequence by the tool, the program provides the UI for users to do the manual change and use these modified information as the training data to process the training, then update the knowledge base of the program as well in order to keep improving the performance of the program in the translation. The program only have limited capability in translating PL/SQL stored procedures because it still can’t mapping the sentences one by one between the source and translation. In the following experiments, we get the conclusion that the translation of the single SQL sentences has the much higher accuracy rate than the translation of the PL/SQL stored procedures. At last, there is a summary of the shortcomings existing in the program and the way to improve the program in the future. |
PDF Full Text Request