STRUCTURED SENTIMENT ANALYSIS MODELING WITH BART BASED ENCODER-DECODER AND BI-LSTM.

• Main Author: Putra Tjandra, Andrianata
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/77876
• Institution: Institut Teknologi Bandung

Sentiment Analysis or SA is one of the most widely used fields in Natural Language Processing or NLP today. SA itself also has a lot of branches such as aspect-based, end2end, targeted, and more. Because of this many branches, it becomes difficult to track the development of SA as whole. To overcome this, a solution is given in the form of Structured Sentiment Analysis which aims to predict the holder, target, expression, and polarity of sentences collectively in a tuple. One of the methods used is the BART-based Encoder-Decoder, but this method has a weakness in handling long sentences and sentences without opinions. Therefore, a solution is proposed to overcome these weaknesses by combining BART with an RNN or Recurrent Neural Network. From a number of alternative RNNs that have been studied, a bi-LSTM was finally selected for use. From the results obtained, the use of bi-LSTM can improve the performance of BART-based Encoder-Decoder.