Improving user experience of a P2P review platform with Machine Learning and NLP

The client’s online marketplace includes over 2,000 categories of software products and services users can leave reviews for. They wanted to help their users quickly and easily find all the details they need about products (such as their advantages and disadvantages).

Additionally, the client wanted to improve the SEO ranking of their marketplace to attract more users and increase marketplace traffic. To achieve this, they wanted to be able to identify specific keywords in user reviews, cluster them, and generate topics.

N-iX helped the client implement a Pros and Cons feature on their marketplace powered by Natural Language Processing (NLP) and Machine Learning (ML). The feature analyzes the advantages and disadvantages of each category of every software product, summarizes them, and determines the topics as the product’s main pros and cons. The keywords used in topic names represent what most users liked and disliked about a product.

We have utilized GPT-4 with topic modeling to analyze user feedback and their replies related to product quality. LangChain is used to handle API requests to the GPT-4 large language model. We have also built and implemented a BERTopic pipeline for topic modeling.

In addition, we have developed a feature that shows the number of user reviews on a particular topic. It provides users with a list of reviews with keywords related to the topic they are interested in. This makes it easier to find a product with a particular attribute.

N-iX engineers have identified the most effective combination of techniques that makes GPT-4 provide consistent answers for multiple similar requests. We have also implemented topic IDs to prevent GPT-4 from making up topics.

Furthermore, we have made optimizations to the prompt to make sure it contains the least possible amount of tokens. This was done to help the client optimize their expenses, since OpenAI API charges for the amount of tokens sent in the prompt and the number of tokens in the output.

Lastly, we have helped the client create a unified solution for clustering differet data distributions. Since each product category contains a different number of keywords, parameters in the cluster algorithms could not be effectively used for all product categories. Hence, our engineers investigated various keyword clustering approaches and set the parameters that can be applied within the algorithm based on data distribution.

N-iX has helped build and implement a Pros and Cons feature powered by Machine Learning and NLP on the client’s marketplace. We have also developed a feature that shows users the number of user reviews based on a particular topic. As a result, the client has benefited in several ways:

• Enhanced user experience with an effective Pros and Cons feature for listed products and services, attracting more users and increasing marketplace traffic;
• Optimized costs associated with OpenAI by making prompt optimizations;
• Improved the marketplace’s SEO ranking by defining keywords from user feedback, clustering them, and generating related topics with the help of ML and generative AI.