Sentiment Analysis of Cooking Oil Price Policy in Indonesia using Word2Vec Feature Extraction and Bidirectional Long Short Term Memory

The policy of the Ministry of Trade of the Republic of Indonesia which sets the highest retail price for cooking oil and then cancels it in early 2022 due to people's complaints about the difficulty of getting cooking oil in the market has received many responses from the public via social media. This study conducts sentiment analysis using Word2Vec and Bidirectional Long Short Term Memory (BLSTM) with the aim of studying the public's response to the removal of the highest retail price policy and effect of both skip-gram (SG) and continuous bag-of-words (CBOW), window size and dimension to the word similarity and accuracy. Data taken via Twitter in March 2022. The data then goes through preprocessing stages, manual labeling by experts, and word embedding using Word2Vec to be further classified using BLSTM. Data labeling produces imbalanced data where the largest is neutral sentiment (53%) followed by negative (27%) and positive (20%). ome of the words such as ‘mafia’, ‘mahal’, ‘langka’, ‘tangkap’, ‘bongkar’, and ‘mendag’ could become the government's attention. The similarity value of the SG model tends to have a higher value than CBOW at a lower word vector dimension. The model that gives the best performance results is the Word2Vec model with 50 dimensions, SG architecture and window size 5. This is shown in the accuracy of 0.71 which is the highest value and only shows slight signs of overfitting. Symptoms of over fitting will be more visible with increasing dimensions.