DEVELOPMENT OF AUTOMATED WEB SERVICE CLASSIFICATION SYSTEM USING NATURAL LANGUAGE DESCRIPTION AND TEXT MINING

• Main Author: Yazid Albisthami, Hafshy
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/74779
• Institution: Institut Teknologi Bandung

The development of web services in the world of service computing continues to evolve, and its application in business sector is becoming more advanced. As technology progresses, many web services are transitioning from SOAP to RESTful. Consequently, the use of structured descriptions like WSDL is being phased out. There are various service descriptions available for RESTful services, such as OpenAPI Specification, which includes both structured and natural language descriptions. Manually performing service discovery based on these structured or natural language descriptions can be complex and time-consuming, particularly when dealing with a large collection of services. To address these challenges, automatic web service classification using service descriptions can streamline service discovery process. Development of an automated web service classification system involves employing text mining methods, including dataset search, preprocessing, building a classification model, and constructing automatic classification system. Dataset used in this study consists of web service descriptions obtained from ProgrammableWeb, categorized into 5, 10, 15, 20 labels. Methods used to build classification model include problem transformation, adapted algorithms, ensemble, LSTM and BERT. Based on the conducted experiments and tests, the best model is label powerset with SVM linear, which falls under problem transformation method. This model exhibits most favorable evaluation metrics and can classify web services in shortest amount of time. Hamming loss values for the 5, 10, 15, and 20 labels are 0.089, 0.065, 0.059, and 0.049, respectively. On average, the number of predictions that can be made in one minute for the 5, 10, 15, and 20 labels are 6,122; 3,496; 2,374; and 1,220; with a total model size of 191.5MB.