Defect product estimation using customer reviews, Amazon use case

Commerce is highly affected by technological improvements, like every other field. Today, all businesses who want to reach end-users, such as manufacturers, retailers, or service providers, can reach their customers quickly over the internet through methods such as e-commerce sites and mobile applications. On the other side, customers now have the opportunity to choose from many options. While deciding, users generally benefit from the comments of other users who have shared the same experience before. In this respect, user comments contain very valuable information. However, on heavily used sites, so many comments accumulate that a person cannot review them one by one. In this study, we focused on a certain feature of the products, namely their defective features, and we propose a method to filter related comments from millions of comments containing the ones that include only defect information. A simple solution may be manually creating a word list related to defects and filtering the comments that contain these words. But manually creating the word list will not be efficient. For this, we trained our own word-embedding model using only the comments of the relevant product groups and created a more efficient list of defect words list by using word similarities using this model. We downloaded a ready-to-use pre-trained word-embedding model and compared it with our own model. We observed that the pre-trained model is more successful in general tasks, while our own model is more successful in creating a word list on a context-specific task.