2024-06-11 –, Maschinenhaus
Exploring and enhancing lexical search and semantic search in practical scenarios, we assess various optimization methods and their varied effects on metrics. The focus is on the integration of semantic search into an established lexical search system, addressing potential challenges and pitfalls.
The AI wave has hit the search community hard, making semantic search a hot topic. As a search engineer, you've probably worked on it too, developing PoCs or maybe even running online A/B tests in production. However, it's not uncommon to conclude that "Semantic search doesn't quite fit our domain," leading us back to the familiar territory of lexical search with its established synonym dictionary and knowledge graph. But why does semantic search often underperform both offline and online?
This talk explores these critical questions, sharing insights from years of trial and error. We'll start with a detailed comparison of lexical and semantic search from various perspectives, including the validity of their evaluation metrics. A quantitative analysis will then highlight their strengths and weaknesses, providing a balanced view of both systems. A key part of our discussion will be how to integrate semantic search into existing lexical search systems. Most importantly, we'll address a crucial question: Is investing in semantic search integration really worth it? Attendees will leave with a clearer understanding of the semantic search landscape and the knowledge to make informed decisions about its implementation in their projects.
Kentaro is a Search ML engineer at Mercari, Japan’s largest C2C marketplace. While Kentaro is working on models, data pipelines, and evaluation for search, he is also exploring methods to effectively combine a complex search system, built on top of existing search technologies, with new technologies.