In this study, we extract main components and attributes, which are the main aspects of Electric Vehicle by analyzing User Experience based on Aspect-Based Sentiment Analysis (ABSA) using machine learning, overcoming the problems accompanying in this ...
In this study, we extract main components and attributes, which are the main aspects of Electric Vehicle by analyzing User Experience based on Aspect-Based Sentiment Analysis (ABSA) using machine learning, overcoming the problems accompanying in this process such as Data Imbalance and insufficient user reviews by making use of non-label data. In addition, we find the contributing factors affecting user’s sentiments by figuring out the relationship between user's sentiment to each aspect extracted and detailed specifications of Electric Vehicle with regression.
Based on the ABSA method, and we perform data collection, data preprocessing, feature engineering, Aspect Extraction, modeling for sentiment analysis, and evaluating user sentiment to each aspect in sequence. For data collection, a total of 5,065 label data, which is evaluated with a 5-point scale by users, was collected from representative car forums. At the same time, in order to overcome the shortage of data and data imbalance, approximately 210,000 items of non-label data are collected from Youtube.com, of which 6,488 items were selected by filtering with limited to the user experience related only. And then, feature engineering is performed with effective embedding methods of distributed representation after data pre-processing. The analysis phase is mainly divided into two processes: Aspect Extraction and Sentiment Analysis. First of all, TextRank and Naïve methods were used as an unsupervised method and an extractive approach for Aspect Extraction. Then, in order to implement a sentiment classification model based on supervised learning with high performance, we built a machine learning model that trains the truncated text composed of one or two sentences at the beginning of a review text with a label and make it improved by means of semi-supervised learning. With the model trained, we are able to perform aspect-wise sentiment analysis by conducting sentiment analysis on the sentence that including the selected aspect term. Further, we find detailed specifications of vehicle that have an influence on user sentiment as contributing factors that affects user’s sentiment.
As a result, 16 categories of main aspects were extracted, eight key EV Components & eight key Human Factor Attributes, of which the users are likely to be positive to “Acceleration, Room, Interior, Power, Safety, Ergonomics, Price, Power” and negative to “Seat, Battery, Charge, Noise, Winter, Ice”. In sentiment analysis, the CNN model showed the highest performance in sentiment classification. Therefore, through semi-supervised learning using CNN, label propagation was performed among non-label data, giving the pseudo label to only the data with a high classification probability more that 80%, resulting in improvement in performance of the CNN model. Lastly, we confirmed the high classification accuracy of the deep learning model for predicting the user’s sentiment of the sentences. In addition, with regard to aspect-wise sentiment analysis, there was a tendency to predict the user’s sentiment similarly between machine learning based and lexicon-based, which showed machine learning based model is robust as much as lexicon-based.
In conclusion, it was shown that more diverse topics and unbiased opinions could be extracted through aspect-wise analysis than review-wise. In addition, we verified that the imbalance problem could be overcome by over-sampling Finally, a more effective UX analysis framework for the products that have not sufficient user reviews was proposed by taking advantage of non-label data with semi-supervised learning.