Validation Optimization using NLP and Machine Learning Techniques

Abstract

Integration and validation is the most vital part before releasing products to customers in Intel. The validation team approves the release based on multiple stages of validation on hardware and software stack. Validation of software stack is performed by executing test cases for each domain. Bugs are raised after execution of test cases on each platform and they arise due to mismatch of product specifications. Hence, similar bugs arise after filing bugs and many issues are closed as duplicates. The main objective is to find the duplicate bugs before filing a new bug. Hence, debug efforts can be reused and optimised. Duplicate bugs are found by searching the bugs being filed by the user in the database of defects using ElasticSearch engine. Proposed scoring algorithms based on driver versions and platform hierarchy are then applied to rank the possible similar bugs. LSTM neural networks are also incorporated to predict duplicate bugs by considering context of the sentence and thereby, trying to improve accuracy. The accuracy after validating the bugs found through ElasticSearch manually is 60 % . The validation accuracy of predicted duplicate bugs through LSTM networks is found to be 80 % , but further investigation is required to improve accuracy of predictive bugs. BERT networks can be used to enhance the context of the sentence for improvements. The accuracy might also improve by searching the bugs in the particular domain defect database through ElasticSearch engine.