Our client, a large Amazon E-commerce seller, who frequently launches new products and seeks continuous improvement for existing ones.

The client recognized the value of their own reviews and those of their competitors to understand customer pain points. However, the manual process of extracting insights from reviews is a significant challenge.

This case study presents a solution that effectively automates the extraction of valuable insights from unstructured text data.

• Scalable Solution: Frequent product launches and iterations demanded an efficient and automated approach to extract valuable information from customer reviews.
• Multilingual Support: The solution required the ability to perform topic extraction and sentiment analysis on unstructured text data across multiple languages.
• Limited Training Data: The absence of training data is an additional challenge in developing an effective solution.

To address these challenges, we opted to utilise large language models, specifically OpenAI's GPT-3.5-Turbo model.

This model, partially trained on Amazon reviews, offered significant advantages.

Benchmarks indicated that GPT-3.5-Turbo outperformed specialised sentiment analysis systems in many cases. As customer reviews are public data, data privacy concerns were minimal, making GPT an ideal choice.

Proof-of-Concept (PoC):

Before developing a full-scale application, we initiated a Proof-of-Concept phase to demonstrate the expected quality of the solution.

Development of the Application:

The implementation of the solution involved the following steps:

1. Data Scraper: We built a data scraper that collects reviews from product listing pages and stores them in a database. Existing reviews are automatically skipped to avoid duplication.
2. Topic Extraction and Sentiment Analysis: Leveraging GPT-3.5-Turbo, we performed topic extraction and sentiment analysis on the collected reviews. The sentiment analysis ranked the sentiment on a scale from -1 to 1.
3. Storing Results: The extracted topics and corresponding sentiment rankings were stored back in the database for further analysis.
4. Visualization Interface: To facilitate easy access to insights, we developed a simple interface that visually presents the results, including the most frequently mentioned issues.

The implementation of this automated review analysis solution yielded the following benefits for our client:

• Improved Decision-making: The client gained better insights into which products to source and how to enhance them, enabling informed decision-making.
• Better Issue Detection: By automating the review analysis process, issues with their own products were identified faster and in a more structured manner.
• Cost Savings and Improved Reporting: In addition to cost savings resulting from the automation, the quality of the generated reports surpassed those produced by the previous method, which involved a combination of virtual assistants and in-house team members.