TITLE RECOMMENDATION FOR INDONESIAN NEWS ARTICLE USING LONG SHORT-TERM MEMORY WITH ATTENTION MECHANISM

• Main Author: Cahyadi, Joshia
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/76269
• Institution: Institut Teknologi Bandung

News articles are one of the most widely circulated sources of information on the internet. The existence of news articles makes people more sensitive toward the situation happening in the real world. Title, in this case, certainly has an important role to give some descriptions of an article. Unfortunately, nowadays many titles are made very interesting, but not following the content of the corresponding news article. This final project focuses on developing a title recommendation program specifically tailored for Indonesian news articles using a deep learning approach. The developed title recommendation program in this final project leverages the sequence-to-sequence architecture with long short-term memory (LSTM) as the basis. Additionally, this architecture also incorporates an attention mechanism to help the model to capture important information inside the news article. There are 301,000 pairs of Indonesian articles and news titles used to train and evaluate the model. Experimental results demonstrate that the designed model in this final project has some abilities to understand the topic of a news article. The model can also generate a title that has some degree of similarity to the content of the corresponding news article. Quantitative evaluation has been done using BERTScore as the evaluation metric between the generated title and the news content used as the input for the model. Overall, the model with an attention mechanism has a relatively better performance than the model without an attention mechanism.