Improved recommendations for scholarly literature using GenAI

Challenge

In modern e-commerce, accurate and personalized recommendations play a central role. For complex products and services, targeted search results and recommendations are crucial for user experience. However, practical implementation often proves challenging. Our client, a leading provider of legal texts in the field of tax law, faced the challenge of optimizing the search functionalities on their website. Due to the complexity of the subject matter, simpler systems did not yield the desired results, leading to a suboptimal user experience.

Approach

Given the complexity of the texts, we decided to use a special language model pre-trained on legal documents. Using this model, individual paragraphs and entire documents were transformed into a semantic representation (known as "embeddings"). These embeddings allow semantic comparisons between user search queries and all available legal texts. Based on these comparisons, a list of relevant results can be presented to users. Additionally, previous search queries or purchases can be used as additional information to personalize the results.

Result

Compared to the previous search, which was based on keyword search, significant improvements in results were realized. The language and domain knowledge of the language model allowed for better consideration of nuances in search queries, resulting in more relevant outcomes. Customer retention was also improved by further personalizing search queries. The continuous learning of the AI model reduces maintenance efforts in the long term.