As parents, we try to select the perfect toys and learning tools by carefully matching child safety with enjoyment; in doing so, we often end up using search engines to find the right pick. However, search engines often provide non-specific results which aren’t satisfactory.
Recognizing this, a team of researchers have devised an AI model named STARK (Semi-structured Retrieval on Textual and Relational Knowledge Bases), designed with the ability to provide specific search results based on the input of both textual and relational information.
To understand this better, let’s consider a scenario. Suppose you want to buy a tricycle for your child that is both fun and safe from the brand Radio Flyer. It’s quite a specific request, right? The STARK model is designed to understand this complex requirement including both textual (‘fun’ and ‘safe’) and the relational (the brand ‘Radio Flyer’) aspects.
To develop such an AI model, the researchers first built semi-structured knowledge bases using public datasets. The created knowledge bases included details about Amazon products, academic papers and authors, and biomedical entities such as diseases, drugs, and genes. These knowledge bases serve as the information repository accessed by STARK to fetch relevant information.
Next, they created an automated pipeline to generate search queries. This pipeline starts by identifying a relational requirement (such as a product belonging to a particular brand) and then extracts relevant information about the product (like a tricycle being fun and safe for children). The information is then transformed into a naturally sounding query which STARK uses to search the information repository.
The researchers also ensured that the generated results are checked using multiple language models to confirm they meet the query requirements. The checked entities are then added to the final response set.
The researchers found that currently available retrieval models showed less accuracy when reasoning over both relational and textual information. Combining traditional vector similarity methods along with language model rerankers like GPT-4 demonstrated the best results. However, there is significant room for improvement, especially in terms of reducing retrieval latency and incorporating strong reasoning abilities.
The STARK model provides a novel benchmark and valuable opportunities for future research in the area of multimodal retrieval challenges. Given its potential to evolve and redefine how search engines operate, the researchers have made STARK’s work open-source, inviting further exploration and development in this arena.
