SENTIMENT CLASSIFICATION OF CORPORATIONS IN NEWS ARTICLES USING WORD EMBEDDING

• Main Author: AODYRA KHAIDIR - NIM : 13513063 , MUHAMMAD
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/23197
• Institution: Institut Teknologi Bandung

Sentiment of company influences an individual or society view of the company. This view affects individual's actions in stock trading transactions. These corporate sentiments can be contained in news articles in several factors related to the company. These factors are internal factors and external factors of a company. Sentiments contained in the news article can be predicted with the sentiment classification. Sentiment classification facilitates the individual or society in determining their actions in stock trading. <br /> <br /> <br /> <br /> The problem of sentiment classification can be solved using supervised learning. To form a model of learning with supervised learning, a feature extraction method is performed to obtain representations of a sentence. Feature extraction methods that can be used are the features of n-gram and word embedding. Related research produces good accuracy using the Support Vector Machine and Convolutional Neural Network. The feature selection can be used to reduce the size of the vector space model so that the machine learning algorithm can run fast. Selection features also eliminate features that are not relevant in the feature extraction process. <br /> <br /> <br /> <br /> The formation of the learning model uses 6,358 sentences with three classes of sentiments, which are positive, negative and neutral. The experiments were experimental n-gram features and word embedding features. In the n-gram feature experiment, experiments were performed on the features of unigram, bigram, trigram, and all three combined. Each experiment was performed with three weights, namely the appearance of a word, term frequency (TF), and TF-IDF. In addition, feature selection is also performed for each of these n-gram experiments. In word embedding experiments, the experiments performed are tuning parameters, average word vector, and Convolutional Neural Network. Methods with word embedding and Convolutional Neural Network algorithms performed with information gain feature selection and word sentiment dictionary yield the best accuracy. The best accuracy produced is 85.729%.