This article details the engineering challenges and solutions for deploying a 3.5 billion‑parameter MoE LLM in Taobao's search relevance pipeline, covering large‑batch scheduling, dynamic load balancing, intra‑batch KV‑Cache reuse, and MoE kernel tuning to meet sub‑second latency requirements.
The article details the development and deployment of LORE, a large generative model that reshapes e‑commerce search relevance by combining knowledge injection, chain‑of‑thought reasoning, and multimodal alignment, achieving simultaneous improvements in user experience and revenue metrics.
This article explains how a leading e‑commerce platform leveraged large language models to overcome traditional search relevance challenges, detailing the iterative workflow, model distillation, performance gains, deployment results, and future directions for smarter, more accurate product search.
The paper introduces SimBert, a later‑fusion model that jointly trains a dual‑tower retrieval component and an auxiliary generation task on the item tower, using a two‑stage pre‑training and fine‑tuning pipeline, which yields a 3.6% relevance boost and reduces bad‑case rates in Xianyu search.
This article presents an in‑depth exploration of an interactive BERT‑based relevance model for health e‑commerce search, detailing the business context, query and product feature extraction, domain‑specific sample generation, model architecture enhancements, offline and online performance gains, and practical deployment through knowledge distillation.
This article presents the QQ Browser search relevance team's experience integrating QQ Browser and Sogou search systems, detailing business overview, relevance system evolution, algorithm architecture, evaluation metrics, deep semantic matching, relevance calibration, and model distillation techniques to improve search relevance performance.
This article presents a comprehensive overview of Alibaba Health's interactive BERT approach for improving relevance in health e‑commerce search, covering business background, model design, domain‑specific data construction, knowledge‑distilled twin‑tower deployment, experimental results, and a detailed Q&A session.
The QQ Browser search relevance team describes a unified, billion‑scale architecture that combines a main and vertical subsystem, a pyramid‑shaped ranking pipeline (recall, coarse, fine), a dedicated GPU‑accelerated relevance service, and hybrid semantic‑matching models (dual‑tower, BERT, matrix fusion) evaluated with offline and online metrics to deliver accurate, fresh, and authoritative results for diverse content and long‑tail queries.
The article details Meituan‑Dianping's search relevance pipeline, describing how multi‑similarity matrix structures, multi‑stage domain‑adaptive training, POI field summarization, and online inference optimizations together improve a BERT‑based relevance model's offline metrics and reduce the BadCase rate in production.
This article examines Hulu's video search engine, highlighting challenges such as ensuring relevance beyond text matching, handling ultra‑short queries on TV remotes, addressing content gaps, and integrating AI‑driven query understanding, retrieval, and ranking to improve user experience.
This article describes how Haodf Online improved its medical question‑answer search by analyzing search challenges, adopting semantic similarity models based on pre‑trained language embeddings, designing contrastive training tasks, and evaluating the resulting increase in click‑through rate and user engagement.
The paper describes Xianyu’s new relevance‑matching pipeline—integrating basic, text‑matching, semantic (BERT‑based dual‑tower), multimodal, and click‑graph features and fusing them with a GBDT model—which boosts search DCG@10 by 6.5 %, query satisfaction by 24 % and click interaction by over 20 % while outlining future enhancements for finer attribute matching and richer structured data.
Baidu’s elastic nearline computing architecture inserts an asynchronous, resource‑adaptive layer between online and offline processing, dynamically harvesting idle CPU, GPU and Kunlun XPU capacity to pre‑compute complex recommendation and search policies, enabling peak‑shifting, valley‑filling, higher timeliness and significant business growth at low cost.
This talk by Zhihu search algorithm engineer Shen Zhan details the evolution of text relevance models from TF‑IDF/BM25 to deep semantic matching and BERT, explains the challenges of deploying BERT at scale, and describes practical knowledge‑distillation techniques that improve both online latency and offline storage while maintaining search quality.
Using a fruit shop analogy, the article explains recall and precision metrics, illustrates their impact on recruitment search, and presents a matrix of design patterns—including cross‑database search, preset search sets, and matching labels—to boost both recall and accuracy while maintaining user experience.
This article introduces the background of search relevance, reviews three common matching model types—representation‑based, interaction‑based, and hybrid—describes their architectures such as DSSM, CDSSM, DRMM and DUET, and presents experimental results of these models on DiDi Food’s search system.
The article outlines DiDi Food’s search relevance challenge, defines semantic matching versus traditional keyword methods, describes the recall‑ranking pipeline, and reviews three families of deep matching models—representation‑based (e.g., DSSM), interaction‑based (e.g., DRMM) and hybrid (e.g., DUET)—including experimental results and a recruitment notice.
Alibaba’s AliCoCo, a large‑scale e‑commerce cognitive concept net, models user needs as graph nodes, linking concepts, primitives, taxonomy and items, and leverages advanced NLP, BiLSTM‑CRF, projection learning and knowledge‑enhanced models to boost search relevance, recommendation diversity, and overall user experience.
The system tackles POI category tagging as a multi‑label classification problem by engineering textual and non‑textual features, mining click‑log and external samples through active learning, and deploying hierarchical and per‑tag deep textCNN models with feature fusion, achieving over 5 % accuracy gain, ten‑fold speedup, and markedly higher precision and coverage that boost map‑search relevance.
This article explains how Elasticsearch computes relevance scores using Lucene's practical scoring formula, term frequency, inverse document frequency, field-length norms, and query normalization, and demonstrates query-time boosting, constant_score, function_score, decay functions, and script_score with practical DSL examples.
At the 2017 QCon Beijing conference, Suning’s Silicon Valley Research Institute director Jim demonstrated how deep‑learning techniques can transform e‑commerce by vectorizing product data for smarter search relevance and by combining AI models with limited labeled data to build conversational chat‑bot platforms that understand user intent.
