| Title |
Classifying Customer’s Written Questions by Topics Using Deep Learning (LSTM) |
| Authors |
나상태(Sang-Tae Na) ; 양광동(Gwang-Dong Yang) ; 신재섭(Jae-Seop Shin) |
| DOI |
https://doi.org/10.5370/KIEE.2019.68.11.1411 |
| Keywords |
LDA; LSTM; Topic Modeling; Confusion Matrix; e-CRM |
| Abstract |
In response to questions or complaints posted by customers on the company's homepage, the response time is an important measure of customer satisfaction. However, the time it takes for a customer to receive an answer includes a time for the article to be selected by the person in charge of the reply, which limits the shortening. In this study, we developed a model in which a machine, not a person, reads the article, classifies the topic, and delivers it to each person in charge of the article. The article posted on the KEPCO homepage used in this study is a short sentence consisting of an average of 49 words. Due to the scarcity of multi-frequency words, it was found that there is a limit in securing a certain level of topic modeling accuracy in unsupervised machine learning like LDA. To overcome this, we labeled topics and let the machine conduct supervised learning. Although there are limitations in improving accuracy because there are articles containing more than two topics in one article, the classification accuracy is secured up to 84% by using LSTM and Baysian Optimization. The result of this study suggests that topic classification is possible for short-term customer questions in specific fields such as the electric power industry. In addition, it is expected that a model will be developed that can provide optimal reference answers for newly received questions when the topic-labeled questions and answers are fully accumulated. |