Context: the company has a mobile app in the App Store. In the reviews of this app, users write their feedback, which is categorized manually. Analysts use this to create tasks for the Development Department.
Decision: a model was created for clustering and searching for new topics in the feedback stream to automate the work of analysts.
Integration into the BI of the customer service Department automation of review categorization:
- Increase accuracy by 15% (reduce human error);
- Reduced analysis time by 85%.
Harmonization of taxonomy:
- Reducing the time to identify a new category by 60%;
- Adding and merging 15% of categories.
Sentiment assessment and trend analysis:
- Identifying growth points when analyzing 40% more reviews;
- Reducing the time to identify critical and reputation-affecting reviews by 70%.
To build the model, we used:
- Reviews of the mobile app from the AppStore, Google Play, and other sources;
- Assessment of feedback provided by the user;
- The original taxonomy of categories of the reviews.
- Model for categorizing reviews by taxonomy and predicting sentiment;
- Recommendations on harmonization of taxonomy;
- The web application aggregates analytics based on reviews.
Customer: Restaurant, Retail
Technology stack: BigARTM, Python, Flask, PyTorch, nltk, gensim.