Research and Development of Financial Contract Text Information Extraction System based on M-BiLSTM and Our-M Models

With the increasing demand of financial industry for efficient information extraction in contract management, how to accurately extract key information from long-form contract text has become an urgent problem to be solved, while traditional methods are often limited by inaccurate information positioning and poor extraction effect in processing long text. Therefore, this paper proposes a financial contract information extraction system that combines multi-layer bidirectional short-duration memory network M-BiLSTM and joint learning model Our-M. In the information positioning stage, multiple BiLSTM models are integrated to solve the problem of sample imbalance, thereby improving the positioning accuracy. In the content extraction stage, the text segmentation and classification tasks are learned jointly by Our-M model to further improve the extraction precision. The experimental results show that the F1 value of M-BiLSTM model in information positioning tasks is improved by 22.6% compared with the traditional BiLSTM, and the F1 value of Our model on open data set WikiSection and actual contract data set is improved by 1% to 2%. Finally, the overall F1 value on the ContractData2W dataset reaches 83.21%, which proves the effectiveness and superiority of the method. Based on this, the developed contract information extraction system has been successfully launched in a bank and achieved remarkable results, which provides strong support for intelligent information extraction of financial contract management.