Semantic Product Search for Matching Structured Product Catalogs in E-Commerce

Summary

The problem with product search is that product unlike other search has different attributes, i.e, dimension, color which can't be represented like a text, which makes it hard to embed a product to latent space. In this paper, the authors have used all the field embedding and aggregate them to generate single vector embedding. In information retrieval, there are typically 2 stages (a) candidate generation & (b) Re-ranking. In this paper, the authors have worked on the former one, so the ranking doesn't matter and recall matters a lot.

Model Architecture:

So each query goes to a transformer block to generate a Qemb and for each document D, there are multiple fields F1,F2...Fn which again goes into the shared encoder to generate F1E,F2E,...FNE, The authors have mean pooled all the token embedding to generate the final embedding. These two embedding are then stacked. By subtracting and multiplication, the authors created two cross-matrix (representing the pairwise connection between Q and F). On the other side, for lexical matching, the authors have created a binary Q×D matrix. All three matrices then goes through a 2 layer neural network to generate a score which is optimized using binary cross entropy loss.

Fields in Product

For each product, the authors have used following fields to express the product:

Annotations

Annotation

« (1) effectiveness of field representations and structured matching »()

Annotation

« (2) effectiveness of adding lexical features to semantic search »()

Annotation

« To be successful, models first need to understand the semantics of each field »()

Annotation

« (1) candidate generation; (2) candidate re-ranking »()

Annotation

« we propose a structured matching module (SMM) that leverages multiple fielded representations to learn a structured matching function. »()

Annotation

« For evaluation, we trained and validated our model using two data sources in the home-improvement domain: (1) internal user click logs; (2) product search relevance (PSR) dataset »()

Annotation

« his method has shown to be effective compared to directly using the vector for [CLS] token or max-pooling »(2)


Related Notes