Behavior-Driven Query Similarity Prediction Based on Pre-Trained Language Models for E-commerce Search (Amazon Science, SIGIR 2023)

Forum topic · 小凯 · 2026-07-05

Summary

This forum post indexes an Amazon Science publication presented at the SIGIR 2023 eCommerce workshop, titled 'Behavior-driven query similarity prediction based on pre-trained language models for e-commerce search.' The work addresses a core problem in e-commerce search: accurately predicting similarity between user queries so that semantically equivalent searches retrieve the same relevant products. Rather than relying solely on query text, the approach incorporates user behavior signals—such as click and purchase patterns—to supervise or augment pre-trained language models, enabling the model to capture intent that raw strings miss (e.g., 'gift for mom' vs. 'mothers day present'). The post provides background on neural information retrieval, outlines the typical pipeline of representation learning, retrieval, and re-ranking, and discusses engineering trade-offs such as latency, cost, data quality, and evaluation gaps between offline metrics and online satisfaction. It also situates the paper within the broader evolution from BM25 and dense retrieval to LLM-based and agentic search systems. Readers should consult the original publication via the linked Amazon Science page for exact experimental results and quantitative benchmarks.

Behavior-Driven Query Similarity Prediction Based on Pre-Trained Language Models for E-commerce Search

Source: Amazon Science — presented at the SIGIR 2023 eCommerce workshop.

Overview

This post indexes an Amazon Science publication from the SIGIR 2023 eCommerce workshop. The work tackles a fundamental problem in e-commerce search: query similarity prediction — determining whether two user queries express the same shopping intent, so that retrieval and ranking systems can return consistent, relevant results.

Motivation

Query similarity is hard to judge from text alone. Users describe the same product in very different ways (brand names, colloquialisms, vague needs like "gift for mom"). The paper's key idea is to make the prediction behavior-driven: use user behavior signals (e.g., clicks, purchases, session co-occurrence) as training signal for pre-trained language models (PLMs), so the models learn intent grounded in what users actually do, not just lexical or purely textual semantics.

Approach Context

The typical pipeline for such systems follows:

1. Input & representation — encode queries (and optionally user context) into dense representations using a pre-trained language model; 2. Training signal — supervised fine-tuning or contrastive learning, with behavior-derived positive/negative query pairs; 3. Serving — the learned similarity model supports query rewriting, retrieval recall, and ranking in the search stack.

Relevance to Search & Recommendation

PLM-based similarity models bridge the gap between lexical retrieval (BM25) and user intent;
Behavior signals from large-scale e-commerce logs provide supervision at a scale human labels cannot;
Engineering constraints (latency, embedding index updates, online/offline metric gaps) remain key deployment considerations.

Related Entries

Behavior Modeling Space Reconstruction for E-Commerce Search
Automated Query-Product Relevance Labeling using Large Language Models
An Interpretable Ensemble of Graph and Language Models for Improving Search

References

Original publication: Behavior-driven query similarity prediction based on pre-trained language models for e-commerce search, SIGIR 2023 eCommerce workshop.

> Note: Quantitative results and detailed experimental setups should be verified against the original publication.