Deep Learning for Text Matching and Ranking at Meituan

Meituan has deployed AI technologies across its services, including search, recommendation, advertising, risk control, scheduling, speech recognition, robotics, and delivery.

To illustrate the construction of its AI “brain”, a series of articles titled “AI in Meituan” are released, and a book “Meituan Machine Learning Practice” is forthcoming.

Background

Recent advances in deep learning have driven research in text processing. Text tasks can be divided into four dimensions: word‑level, sentence‑level, document‑level, and system‑level applications.

Word‑level: sequence labeling with CRF, Bi‑LSTM+CRF, etc.

Sentence‑level: parsers, Shift‑Reduce, Seq2Seq generation.

Document‑level: sentiment analysis with CNN, reading comprehension with RNN.

System‑level: information retrieval, machine translation, dialogue systems.

Deep‑learning‑based Text Matching

Text matching computes similarity between two pieces of text and is used in search, advertising, suggestion, deduplication, etc.

Traditional vector‑space models represent documents as high‑dimensional TF‑IDF vectors and use similarity measures such as Cosine, BM25, etc.

Matrix‑factorization (LSA) reduces dimensionality and captures latent semantics, but struggles with polysemy.

Probabilistic topic models (pLSA, LDA) introduce latent topics to improve semantic representation, with LDA employing Dirichlet priors and inference via Variational EM or Gibbs sampling.

Since 2013, deep learning models such as Word2Vec (CBOW, Skip‑Gram), DSSM, CDSSM, and LSTM‑DSSM have become dominant for semantic matching. Meituan’s ClickNet framework combines DNN/CNN/RNN encoders with business‑specific features to predict click similarity.

Deep‑learning‑based Ranking Models

Ranking selects a small subset of candidates for display in search, recommendation, or advertising. Three classic paradigms are Pointwise, Pairwise, and Listwise.

Pairwise RankNet (2005) uses a neural network to predict pairwise preferences; later extensions include LambdaRank.

Wide&Deep (Google) jointly learns from sparse “wide” features and dense “deep” embeddings.

YouTube DNN ranking model feeds heterogeneous features (user, video, context) through embeddings and fully‑connected layers.

Meituan adapts these architectures (ClickNet‑v1, ClickNet‑v2) to its O2O scenario, emphasizing business features such as user location, merchant rating, and click‑weighting.

Experimental results show ClickNet outperforms linear models and XGBoost on classification tasks (e.g., Higgs dataset) and improves CTR and revenue in production.

Conclusion

Deep learning’s strong fitting ability and reduced reliance on handcrafted features have made it pervasive in text‑related tasks at Meituan, from semantic matching to ranking, with ongoing research in sentiment analysis, dialogue, summarization, and keyword generation.