The Research On Information Extraction And Computer-aided Translation For SWIFT Messages Generation

SWIFT(Society for Worldwide Interbank Financial Telecommunications)is a non-profit international interbank cooperation organization,it runs a world-class financial messaging network for banks and other financial institutions in the world,so as to provide fast,accurate and excellent services to international financial business.Shenzhen securities information Co.Ltd joined SWIFT in 2008,provided Corporate Action Message for SWIFT and other members.SWIFT message generation processing is to extracting data of Corporation Action from shareholders' meeting announcements of the listing corporation,and then the data is translated into English,and finally filled SWIFT template to generate SWIFT message.At present,translation and extraction are dependented on labor totally,low efficiency,data consistency is difficult to guarantee,for these problems,The paper mainly studies the information extraction and aided translation methods for SWIFT message generation.The study is conducted through the following steps: 1.A data extraction method is designed based on classification according to the features of the texts of shareholders' meeting announcements and based on in-depth analysis on the contents of shareholders' meeting announcements.According to the method,the irrelevant text photograph is firstly to be droped,and then random forest classification model is used to obtain the attributes and values paragraphs of meeting information in announcements.Then,regular expression matching is applied to attributes-values paragraphs to obtain the attributes-values.1000 texts of shareholders' meeting announcements in 2014 and 2015 as experiment samples in this paper,as a result.A final score of 0.92 is achieved in the automatic extraction of related data of shareholders' meeting announcements by means of the system and method studied in the paper.2.An aided translation method is designed based on named entity recognition and text similarity according to the features of proposals and addresses of shareholders' meeting announcements.According to the method,the entities in proposals such as institutions,projects,numbers and names are obtained by Conditional Random Fields,and then entities are abstracted and similarity model is constructed with the replaced texts,and finally the most similar translation is selected as the recommended translation for test proposals,then replaced the name and numbers and thus to be the final translation result.With 660251 pieces of proposals in shareholders' meetings in 2014 and 2015 taken as corpus,the modified BLEU* score is reaching 0.83,all matching score is 0.69.A tailored information extraction and aided translation system for SWIFT message generation is constructed by taking advantage of the above-mentioned technonlogy.The system is now proved to be successful in the auotomatic and visual generation of SWIFT messages through production application,which not only contributes greatly to the acceleration of message generation and the reduction of manual intervention and labor costs,but also ensures the consistency of SWIFT messages.