Why Information Structure Matters in AI
In our recent experiments with knowledge bases, we have discovered the importance of how information is structured. This directly impacts the context and overall meaning, as well as the accuracy of AI language models (LLMs) in answering questions and reducing errors.
To help you understand this concept better, we have created a few experiments that you can test yourself. But first, let’s explore why information structure matters.
Imagine walking into a library. In one room, there are books scattered everywhere with mixed-up titles and no order. In another room, the books are organized by subject, author, and publication date. Which room do you think will help you find the book you want more efficiently?
LLMs are similar to libraries. The way knowledge is presented and organized not only affects their understanding but also their output. It’s important to consider the organization of informational taxonomies and hierarchies.
This includes elements like URL structures, folders, and interrelationships between information. By defining the proper context and highlighting what’s critical, we can enhance the performance of LLMs.
Delving deeper, we need to look at the organization within a document. This involves examining the structure, semantics, formatting, and summaries of individual pieces of information.
Now, let’s dive into our findings on the first aspect: informational hierarchy.
At its core, informational hierarchy is all about context. Whether it’s a URL on a website or the structure of folders within a system, hierarchies set the scene and help LLMs understand the importance and relevance of different data points.
For example, consider these URLs:
– ChatbotConferences.com/conferences/2019/nyc suggests multiple events in different cities.
– ChatbotConferences.com/new-york-city offers just a city, lacking context.
– ChatbotConferences.com/nyc/2019 indicates multiple events in NYC but lacks broader context.
To further understand the significance of hierarchies, we conducted a test using two chatbots with different hierarchy organizations:
– Bot 1: Trained on multiple pages, each representing a distinct event and year.
– Bot 2: Trained on a consolidated page that combines all the agendas.
Want to see the results? Test our ‘Good Bot’ and ‘Bad Bot’ here.
Moving forward, it’s important to remember that structure is just the beginning. If you’re interested in building bots using knowledge bases, we have exciting news! Join us in our upcoming Live Workshops at the Chatbot Conference, where you can learn how to build bots and get certified.
Don’t miss out on this opportunity. Dive in and secure your place now.
Cheers,
Stefan