Analyzing and responding system
for reviews

Business Challenge

One of the basic human needs is emotional contentment and feeling that someone’s there to listen and help them. Companies that develop apps and games usually have an internal customer service department. And if there are a few hundred applications to support across a dozen on different platforms? It would take a huge amount of time to process feedback and requests from unsatisfied users, who leave bad reviews and ratings on your apps or games.

Quick feedback to the arising issues contributes both to the app performance and the audience satisfaction with your product. What’s more important, user satisfaction affects the app ranking, and thus installation figures.

Solution Overview

The Qudata team developed a neural network-based system that allows you to quickly and efficiently process user feedback. The implementation of the system boosted the efficiency of the customer support and sales departments, helping to maintain high app store ratings.

The solution architecture has four main stages:

1. Retrieving reviews from Google Play

2. Interpretation of intents and a comprehensive analysis of the received feedback in order to form a response

3. Response monitoring by a customer support manager

4. Sending a response to Google Play

The first step of the solution development was to create a neural network to identify the review intent. For this purpose, we used a recurrent neural network with the Attention mechanism. The network learned from massive feedback data.

The developed system checks reviews on Google Play regularly, depending on the user activity settings. Each review is processed according to its subject matter. The intellectual core of the solution holds 50+ different meanings and moods that can be reflected in the user's message.

The solution has shown good results, with over 90% of reviews successfully processed without manager involvement. For non-standard cases which imply technical issues, there is a notification system with tasks sent directly to the technical support. The solution is integrated with the project tracking system used by the client, so that no review is left unaddressed. After the launch of the service, the rating of supported applications keeps above 4.2 stars.

Technological Details

The project was made with Python. As artificial intelligence, we used the open source RASA platform trained on our dataset. Reply to reviews was implemented using the Google Play Developer API.

The project architecture is universal and can be adapted to different e-commerce sites, app stores and platforms. Additional training of the neural network helps to customize the system for various businesses:

  • online marketplace or retail service;
  • tourist agency;
  • gaming website;
  • site of a cafe or delivery service;
  • educational project;
  • app pages on the AppStore and GooglePlay;
  • any local business acquiring user reviews.

The system supports English and Russian. Modularity provides for the integration with different support systems (Zendesk, JIRA Service Desk, Helprace, Kayako, osTicket, OTRS etc.)

Technology Stack

Python

Python

Rasa

RASA

Python

Python

Rasa

RASA