Meituan SIGIR2020 Workshop: MT‑BERT, KDD Cup Solutions, and Knowledge Graph Applications

ACM SIGIR 2020, the premier international conference on information retrieval, was held online from July 25‑30, 2020 due to the pandemic.

On the morning of July 29 (Beijing time), Meituan organized a workshop at SIGIR2020 featuring three speakers—Wang Jingang, Hu Ke, and Hu Maodi—who presented technical advances from Meituan’s AI platform.

Talk 1 – MT‑BERT: Best Practices of Pre‑trained Language Models at Meituan

Wang introduced MT‑BERT, a BERT‑style model pre‑trained on massive Meituan business corpora. After a brief review of the BERT evolution, he described how MT‑BERT is applied to search intent recognition, fine‑grained sentiment analysis, text classification, and relevance ranking. The model is trained with the in‑house AFO framework on YARN, leveraging Horovod for distributed training and mixed‑precision (FP16/FP32) to achieve a 2.2× speedup. Knowledge‑aware masking incorporates entities from Meituan’s large‑scale “Meituan Brain” knowledge graph, yielding noticeable gains on both generic and business benchmarks. To meet latency requirements, two compression strategies were explored: layer pruning (which preserves performance for short‑text classification) and knowledge distillation (reducing a 12‑layer BERT to a 3‑layer model with 21 M parameters, retaining 98 % of the original performance while delivering a 4.5× inference speedup). The MT‑BERT‑based multi‑task model also achieved state‑of‑the‑art results on the AI Challenger 2018 fine‑grained sentiment dataset and has been deployed in Meituan’s review selection service.

Talk 2 – KDD Cup 2020 Winning Solution and Its Applications in Meituan Search Ads

Hu Ke discussed how Meituan tackled data bias (position bias and selection bias) and applied graph learning and multimodal learning to search advertising. The team introduced a Beta‑distribution‑based exploration algorithm, a negative‑sample generation method, and a multi‑stage training pipeline. Graph learning was used for both trigger and CTR estimation modules, integrating visual features of ads with user‑query semantics. The same techniques powered three KDD Cup tracks: (1) Debiasing—re‑designing the ranking pipeline to generate samples that reflect the true candidate distribution; (2) AutoGraph—an automated graph‑neural‑network search framework that combined spectral and spatial GNNs, achieving competitive results on heterogeneous graph tasks; (3) Multimodalities Recall—using a fusion‑layer architecture and knowledge‑distilled training to improve multimodal matching. Additionally, the team presented TABLE (Task‑Adaptive BERT‑based Listwise Ranking), which ranked first on the MS MARCO document ranking leaderboard, breaking the 0.4 MRR barrier.

Talk 3 – Knowledge Graph and Its Applications in Meituan Waimai

Hu Maodi described the motivation, construction, and deployment of a large‑scale knowledge graph for Meituan’s food‑delivery platform. Because menu items are often non‑standardized, a hierarchical category system and a ten‑million‑scale dish‑name vocabulary were built using topic modeling, multimodal cues, and pre‑trained language models. Entity and relation extraction pipelines (multimodal, multi‑task) were employed to populate the graph with attributes such as ingredients, flavors, and cooking methods. Three concrete applications were highlighted: (1) Package recommendation—using a pointer‑network Seq2Seq model with coverage and hard/soft constraints to generate real‑time, personalized meal combos; (2) Recommendation‑phrase extraction—leveraging a BERT‑based QA model to label BIO tags for sentiment‑rich snippets, followed by text generation; (3) KG‑enhanced recommendation—building heterogeneous graphs that combine behavior and knowledge edges, then applying reinforcement learning (e.g., AC) to produce diverse, explainable recommendations. Online A/B tests showed significant lifts in click‑through and conversion rates.

Overall, the workshop showcased Meituan’s end‑to‑end pipeline from large‑scale pre‑training and model compression to bias mitigation, graph‑enhanced learning, and real‑world deployment in search, advertising, and food‑delivery services.