ADVERSARIAL TRAINING FOR CATEGORIZATION ASPECTS ON MULTILABEL REVIEW TEXT

• Main Author: Abdurrasyid, Azhar
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/64103
• Institution: Institut Teknologi Bandung

ABSTRA CT ADVERSARIAL TRAINING FOR CATEGORIZATION ASPECTS ON MULTILABEL REVIEW TEXT By Azhar Abdurasyid NIM : 13517097 Adversarial Training (AT) is a learning technique that uses adversarial example at the training stage so that the model will be able to handle such input data well. An adversarial example is an input that has been added with small disturbance so that when the model accepts the input, the model will fail to make correct predictions. In addition to being resistant to adversarial examples, AT is believed to be able to generalize the model and prevent overfit because AT can be seen as a regularization technique that introduces different data. One of the challenges encountered when trying to adapt AT techniques for natural language processing tasks is the need for continuous input data. This Final Project focuses on to adapt the AT technique to the task of categorizing multilabel aspects in Indonesian Language. To get continuous input data on text data, this Final Project approaches using input data in the form of word embedding. In addition, this Final Project also adopts the CNN feature extraction topology in the previous aspect categorization research to obtain a continuous feature vector that can be used as input data for the AT model. In this way, two AT models will be built, namely the word embedding AT model and the feature-map AT model. Then to solve the problem of multilabel classification, this final project will compare the use of binary relevance and classifier chain strategies. The test results show that the AT word embedding and AT feature-map models have a better F1-macro performance than the baseline with a sequential score of 0.9253 and 0.9234 in the first test data and 0.8531 and 0.8505 in the second test data. The AT feature-map model succeeded in increasing the baseline average F1-score performance on the adversarial test data with an increase of 0.3291. Meanwhile, the AT word embedding model did not succeed in increasing the baseline average F1-score performance on the adversarial test data. Keywords: adversarial training, adversarial example, aspect categorization, multilabel classification, binary relevance, classifier chain