Abstract
Many companies try to analyze and utilize feedback on services. This can be used for improving service quality or marketing. Until now, most natural language processing studies have attempted to analyze emotions divided into positive, negative and neutral. However, in this work, specific negative reasons are extracted and classified. The dataset is a standard dataset from kaggle that uses tweet data for U.S. airline services. Tweets categorized as negative are labeled with 10 categories of negative reasons. The dataset was divided into train, validation, and test 8:1:1. The learning and classification process was largely divided into two stages. The first is to convert words and sentences into vector values. It is compared and analyzed using Doc2Vec and BERT (Bidirectional Encoder Representations from Transformers) models for embedding and vectorization. The second is to learn and classify sentences transformed into vectors by matching them with 10 negative reason classes. During this learning process, I converted the negative reason into a sentence and attached it to the back of the original text and made new data.I then used BERT's Next Sentence Prediction technique to allow further learning to be performed. This method was able to improve classification accuracy. For each dataset and classification method, metrics were computed, visualized, and compared.