Agrupamento de bugs de software a partir do modelo de linguagem pré-treinado BERT e métodos de clustering

Abstract

The need of software maintenance after its implementation, even the user’s difficulties in using a software application is a reality to this day. A user’s problem or software bug, it is a user’s relate that a particular procedure of the system didn’t work as expected. Some times the related fact concerns to a problem to be solved (e.g. of implementation, configuration of software, lack of users expertise), which have been occurred and had its solution documented. This fact facilitates and speeds up the service from the support team, being not necessary to send the problem to others enterprise departments or spend a lot of time in its solution. Therefore, the present work aims to classify these software problems based in text of users requirements to the support department, in order to find potential solutions rapidly. For this it was created a data base of software bugs in Portuguese, on which the PLN methods were applied to perform the pre-processing and the textual representation of texts of bugs. The textual representation occurred by applying the contextualized pre-trained model BERT in its BERT_base version. The representation of the texts was used in the application of the clustering method DBSCAN, in order to classify them. After performing tests with different hyperparameter values and dimensionality reduction methods, the results revealed that there is no underlying structure from the applied data representation. One hypothesis for the result achieved is that the pre-trained model applied is not able to adequately represent the context of bugs texts in Portuguese. Future works are proposed seeking to find more effective methods for the proposed objective.