Overview of Meituan Technical Team Papers Featured at ACM SIGIR 2022 and Related Works

In 2022, Meituan's technical team had multiple papers accepted at ACM SIGIR 2022, covering opinion tag generation, cross‑domain sentiment classification, dialogue summarization transfer, cross‑domain retrieval, click‑through‑rate (CTR) prediction, dialogue topic segmentation, and other AI‑related topics. This article selects ten representative papers, provides brief introductions, and includes download links for each.

SIGIR is a top‑tier international conference on information retrieval (CCF‑A). The 45th ACM SIGIR Conference (July 11‑15, 2022, Madrid, Spain) received 794 long‑paper submissions (161 accepted, ~20% acceptance) and 667 short‑paper submissions (165 accepted, ~24.7% acceptance).

Paper 01: Personalized Abstractive Opinion Tagging

Download: https://dl.acm.org/doi/pdf (Full Paper)

Authors: Zhao Mengxue (Meituan), Yang Yang (Meituan), Li Miao (Meituan), Wang Jingang (Meituan), Wu Wei (Meituan), Ren Pengjie (Shandong University), Maarten de Rijke (University of Amsterdam), Ren Zhaochun (Shandong University)

Abstract: Opinion tags are short textual sequences summarizing user feelings about a product or service, often consisting of short sentences about specific aspects. Existing opinion tag order reflects generic popularity and ignores personalized preferences. This paper proposes a personalized opinion tag generation framework (POT) that extracts key product information from reviews, tracks explicit and implicit user preferences, and determines the order of key information based on user interest. A hierarchical heterogeneous graph jointly models users, products, aspect tags, and words, while behavior graphs capture multi‑type user actions. Experiments on real Meituan data show strong performance on generation and ranking metrics.

Paper 02: Graph Adaptive Semantic Transfer for Cross‑domain Sentiment Classification

Download: https://arxiv.org/pdf (Full Paper)

Authors: Zhang Kai (Meituan), Liu Qi (USTC), Huang Zhenya (USTC), Zhang Mengdi (Meituan), Zhang Kun (Hefei University of Technology), Cheng Mingyue (USTC), Wu Wei (Meituan), Chen Enhong (USTC)

Abstract: Cross‑domain sentiment classification (CDSC) aims to predict sentiment in an unlabeled target domain using transferable semantics learned from a source domain. Existing methods focus on sentence‑level modeling and ignore rich domain‑invariant semantic information in graph structures (e.g., POS tags, dependency relations). This work introduces Graph Adaptive Semantic Transfer (GAST), which includes a POS‑Transformer to extract serialized semantic features, a Hybrid‑GAT module to generate syntax‑based universal semantics, and an Integrated Adaptive Strategy (IDS) to jointly train the modules. Experiments on four public datasets demonstrate GAST’s superiority over state‑of‑the‑art models.

Paper 03: ADPL: Adversarial Prompt‑based Domain Adaptation for Dialogue Summarization with Knowledge Disentanglement

Download: https://dl.acm.org/doi/pdf (Full Paper)

Authors: Zhao Lulü (Beijing University of Posts and Telecommunications), Zheng Fuja (BUPT), Zeng Weihe (BUPT), He Keqing (Meituan), Geng Ruotong (BUPT), Jiang Haixing (BUPT), Wu Wei (Meituan), Xu Weiran (BUPT)

Abstract: Domain adaptation for dialogue summarization seeks to transfer labeled source‑domain data to unlabeled or low‑resource target domains. Existing methods rely on large‑scale pre‑training, which is costly and data‑intensive. This paper proposes ADPL, a lightweight disentangled knowledge transfer method that uses prompt learning. Three prompts—Domain‑Invariant Prompt (DIP), Domain‑Specific Prompt (DSP), and Task‑Oriented Prompt (TOP)—capture shared knowledge, domain‑specific knowledge, and fluent generation respectively. Only prompt‑related parameters are updated during training, reducing memory requirements. Experiments on QMSum and TODSum show consistent improvements.

Paper 04: Structure‑Aware Semantic‑Aligned Network for Universal Cross‑Domain Retrieval

Download: https://dl.acm.org/doi/pdf (Full Paper)

Authors: Tian Jialin (Meituan), Xu Xing (University of Electronic Science and Technology of China), Wang Kai (UESTC), Cao Zuo (Meituan), Cai Xunliang (Meituan), Shen Fengtao (UESTC)

Abstract: Universal Cross‑Domain Retrieval (UCDR) aims to align and retrieve multi‑domain image representations based on content. Traditional CNN‑based methods struggle with global structure modeling. This work proposes a Vision‑Transformer‑based Structure‑Aware Semantic‑Aligned Network that integrates a self‑supervised ViT with fine‑tuning, aligns multi‑domain representations via learnable class prototypes in hyperspherical space, and prevents forgetting of global structure through soft‑label alignment. Experiments demonstrate significant gains over existing methods.

Paper 05: Multimodal Disentanglement Variational Autoencoders for Zero‑Shot Cross‑Modal Retrieval

Download: https://dl.acm.org/doi/pdf (Full Paper)

Authors: Tian Jialin (Meituan), Wang Kai (UESTC), Xu Xing (UESTC), Cao Zuo (Meituan), Shen Fumin (UESTC), Shen Hengtao (UESTC)

Abstract: Zero‑Shot Cross‑Modal Retrieval (ZS‑CMR) requires retrieving unseen classes across modalities. Existing methods rely on generative models with external semantic embeddings, overlooking reconstruction effects. This paper introduces MDVAE, comprising modality‑specific Disentangled VAEs (DVAE) and a Fusion‑Swap VAE (FVAE). DVAE separates modality‑invariant and modality‑specific features; FVAE fuses and swaps information without extra semantics. A novel counter‑intuitive cross‑reconstruction scheme enhances invariant feature richness. Experiments on image‑text and image‑sketch benchmarks achieve new state‑of‑the‑art results.

Paper 06: Co‑clustering Interactions via Attentive Hypergraph Neural Network

Download: https://dl.acm.org/doi/pdf (Full Paper)

Authors: Yang Tianchi (BUPT), Yang Cheng (BUPT), Zhang Luhao (Meituan), Shi Chuan (BUPT), Hu Maodi (Meituan), Liu Huai‑jun (Meituan), Li Tao (Meituan), Wang Dong (Meituan)

Abstract: Interaction data (e.g., user‑merchant clicks) are increasingly abundant, prompting clustering methods to discover interaction patterns. Existing approaches model objects and pairwise relations as graph nodes/edges, but only capture partial information, either by decomposing interactions into pairwise sub‑interactions or focusing on specific object types. This work proposes CIAH, an Attentive Hypergraph Neural Network that models full interactions as hyperedges encompassing user attributes, merchant attributes, item attributes, and spatio‑temporal attributes. An attention mechanism selects important attributes for interpretable clustering, and a significance‑guided consistency loss aligns attention with true importance. Experiments on public and Meituan datasets show CIAH outperforms state‑of‑the‑art clustering methods.

Paper 07: DisenCTR: Dynamic Graph‑based Disentangled Representation for Click‑Through Rate Prediction

Download: https://dl.acm.org/doi/pdf (Short Paper)

Authors: Wang Yifan (Peking University), Qin Fang (Meituan), Sun Fang (Meituan), Zhang Bo (Meituan), Hou Xuyang (Meituan), Hu Ke (Meituan), Cheng Jia (Meituan), Lei Jun (Peking University)

Abstract: CTR prediction is crucial for recommendation and advertising. Existing models often use single‑item behavior sequences, ignoring the competition among items on a page and missing page‑level feedback such as scrolling or abandonment. DisenCTR introduces a dynamic‑graph‑based disentangled representation framework that captures multiple user interests via dynamic routing on user‑item sub‑graphs and models self‑exciting effects with a Mixture of Hawkes Processes. Experiments on public and Meituan datasets demonstrate significant performance gains.

Paper 08: Hybrid CNN Based Attention with Category Prior for User Image Behavior Modeling

Download: https://arxiv.org/pdf (Short Paper)

Authors: Chen Xin (Meituan), Tang Qingtao (Meituan), Hu Ke (Meituan), Xu Yue (Meituan), Qiu Shihang (Hong Kong University of Science and Technology), Cheng Jia (Meituan), Lei Jun (Meituan)

Abstract: In recommendation ads, POI images influence click behavior. Existing pipelines use a two‑stage approach: a pre‑trained CNN extracts image embeddings, then the CTR model jointly trains with these embeddings, which is sub‑optimal and lacks category priors. This paper proposes a Hybrid CNN (Fixed‑CNN + Trainable‑CNN) architecture (HCCM). The shallow CNN is frozen with ImageNet weights, while the deep CNN is trained jointly with the CTR model. An image‑semantic attention combines candidate images with user preferences, and a channel‑attention mechanism injects category priors into feature maps. Experiments on Meituan’s in‑store recommendation ads show significant improvements across multiple scenarios.

Paper 09: Dialogue Topic Segmentation via Parallel Extraction Network with Neighbor Smoothing

Download: https://dl.acm.org/doi/pdf (Short Paper)

Authors: Xia Jinxiong (Meituan), Liu Cao (Meituan), Chen Jiansong (Meituan), Li Yucheng (Meituan), Yang Fan (Meituan), Cai Xunliang (Meituan), Wan Guanglu (Meituan), Wang Houfeng (Peking University)

Abstract: Dialogue topic segmentation divides a conversation into predefined topical segments. Existing two‑stage methods first split text then label segments, often focusing only on local context and ignoring dependencies between segments, while also suffering from boundary ambiguity and label noise. This work proposes PEN‑NS, a Parallel Extraction Network with Neighbor Smoothing. PEN extracts segment candidates in parallel, optimizes a bipartite matching cost to capture inter‑segment dependencies, and applies neighbor smoothing to mitigate noise and fuzzy boundaries. Experiments on dialogue‑ and document‑based segmentation datasets show PEN‑NS outperforms current state‑of‑the‑art models.

Paper 10: Deep Page‑Level Interest Network in Reinforcement Learning for Ads Allocation

Download: https://dl.acm.org/doi/pdf (Short Paper)

Authors: Liao Guogang (Meituan), Shi Xiaowen (Meituan), Wang Ze (Meituan), Wu Xiaoxu (Meituan), Zhang Chuheng (Meituan intern), Wang Yongkang (Meituan), Wang Xingxing (Meituan)

Abstract: In feed‑stream scenarios, a user’s page‑level behavior is influenced by multiple items competing for attention, making single‑item interest modeling insufficient. This paper introduces DPIN, a Deep Page‑Level Interest Network built on reinforcement learning. DPIN constructs page‑level sequences from historical behavior, employs a self‑attention layer to model intra‑page competition, incorporates page‑level negative feedback (scrolling, abandonment) with denoising, extracts multi‑scale local vision via varied convolution kernels to capture user receptive‑field differences, and matches page‑level historical sequences with candidate sequences. Offline experiments and large‑scale online deployment in Meituan’s food‑delivery platform show substantial performance gains.

These ten papers illustrate Meituan’s collaborative research with universities and research institutes across topics such as opinion tagging, cross‑domain sentiment classification, domain adaptation, cross‑domain retrieval, CTR prediction, and dialogue segmentation, providing valuable insights for the research community.