Total 596 words used.
Karaca is Turkey's leading brand serving in the tableware, kitchen and small electrical appliances categories. It also provides services in the global arena with its more than 300 stores and approximately 4000 employees. At the same time, it gets 1/3 of its turnover from its digital channels. It provides service in Germany, England and France both with its stores and digital platforms. It also sells products on marketplaces in many countries, including America.
Customers share their experiences after shopping through various channels. This feedback, sometimes on social media, sometimes on product comments, and sometimes on a complaint filed with the call center, includes positive or negative thoughts of the customer about the shopping experience or product features. The process of transforming the text, which is included in these feedbacks and shows the development areas for both the product and the company, into information begins with the collection of data. Customer comments on purchases made from the Karaca mobile application, comments made on domestic and international marketplace sites, records created through the call center, return reasons for customers returning their products, and social media. These shares are collected in the database. All collected feedback is firstly sentiment analysis done by an artificial intelligence that classifies the positive or negative state of the experience conveyed by the customer. While doing this analysis, it is classified by an artificial intelligence that can classify Turkish or English according to the feedback language. Correction of spelling mistakes made before this classification with normalization is applied. After negative comments, it is classified according to 6 different complaint categories with an artificial intelligence.During the preparation of this classification artificial intelligence, various machine learning and deep learning models were tried and the results of 56 different scenarios were examined.The models were first trained using the default parameters with these matrices. During this training process, 527,362 positive feedbacks and 276,346 negative feedbacks were used. During the evaluation of deep learning models, LSTM, RNN and BERT models were tested. Embedding processes have been done in order to give words as input to deep learning models. Two different approaches were applied during the embedding trials. In the first approach, the Embedding layer of a pre-trained BERT model is used, while in the second approach, the model's own Embedding layer is created. Deep learning outputs include various analyzes such as the trend of product reviews, comments received by products, product sales, and complaint analyzes for suppliers from which these products are purchased, within a decision support system. This report is shared with the entire company. Customers’ complaints include both a negative customer experience and point to improvement points for the company. Thanks to this project we developed;
· One of the biggest outputs of this project ,quality increase, also decrease in return rates. While decreasing the return rates, an additional income of 7 Million TL per month , 84M TL per year can be provided.
· Since the whole process is carried out by artificial intelligence, the workload required to classify these feedbacks is reduced. An average of 30.000 weekly feedback data is processed automatically. 5 people weekly workload is improved. The annual productivity of this improvement was 1.8 Million TL.
· All feedbacks are collected, processed and reported on a single platform.
· It is ensured that the products are defined not only according to the descriptions written on them, but also with their experiences in the eyes of the customer.
· It is ensured that the brand perception in the eyes of the customer is transformed into a measurable KPI.
Nomination Title: Sentimentality
