INSPIRE: Intent-aware Neural Sponsored Product Retrieval for E-commerce

In the context of Walmart's massive e-commerce grocery platform, search queries are often short, ambiguous, and underspecified. Standard retrieval systems frequently fail to align these queries with the correct products, particularly when user intent is implicit (e.g., a user searching for 'chickpea pasta' implicitly requiring a gluten-free product). The authors seek to improve sponsored search relevance by explicitly modeling and incorporating these latent intent signals into the retrieval process.

INSPIRE (Intent-aware Neural Sponsored Product Retrieval for E-commerce) addresses this by treating intent as a set of structured, multi-dimensional attributes. The framework operates in three main stages:

1. Weakly Supervised Annotation: The authors use an ensemble of large teacher LLMs (Gemma, LLaMA, Qwen) to extract structured intent attributes (brand, flavor, dietary preference, cuisine, etc.) from product titles and descriptions. They apply cross-model consensus to filter out noise and ensure high-confidence labels.
2. Model Distillation: These high-confidence annotations are used to train a lightweight student model via LoRA-based supervised fine-tuning. This allows the system to predict intent attributes at scale with low latency.
3. Intent-Augmented Retrieval: The predicted intent attributes are integrated into a bi-encoder framework. By incorporating these structured signals into both query and product representations, the system achieves more precise matching than lexical or semantic similarity alone.

For e-commerce platforms, small mismatches between user intent and retrieved products lead to significant losses in engagement and revenue. In sponsored search, where ad slots are limited and competition is high, precise relevance is critical. By moving beyond surface-level text matching to understand latent attributes like dietary constraints or specific use-cases, INSPIRE helps ensure that the right products are surfaced to the right users, ultimately improving the return on ad spend for advertisers and the overall shopping experience for customers.