As an important carrier for recording the details of case in the field of criminal investigation, case records contain address of the crime scene which is crucial to the case analysis. This paper uses Lattice LSTM to extract address from case records which combines the advantages of the character-based and word-based models. It can exploit word and word sequence information without performing Chinese word segmentation. For the purpose of optimizing the model, this paper proposes a word vector template which is generated by address repository of Shanghai to pretrain more accurate and more controllable word embeddings. The experiment shows that the model achieves the performance with the precision rate of 96.49%, recall rate of 94.83% and F1 score of 95.65%.