Query Intent Recognition in Enterprise Search: Knowledge‑Enhanced and Pretrained Model Approaches

Background

Enterprise digitalization relies on AI, big data, and cloud computing to transform business and management processes. Alibaba’s internal search platform aggregates content from dozens of sites (DingTalk docs, Yuque, ATA, etc.) and serves over 140 QPS. A unified search engine is needed to avoid information silos and improve relevance.

The search architecture includes a Query Processing (QP) service deployed on the DII platform, which performs tokenization, spelling correction, term weighting, query expansion, and intent recognition before the Ha3 engine performs recall and ranking.

Work Sharing – Case 1: Internal Assistant (内外小蜜)

The assistant uses Alibaba’s DAMO‑Lab Cloud‑XiaoMi QA engine, supporting FAQ, multi‑turn task‑oriented, and knowledge‑graph QA. Intent recognition classifies short queries (most under 10 tokens) into business lines using knowledge‑enhanced short‑text classification.

Knowledge enhancement leverages >6,000 internal knowledge cards and similar historical queries. A dual‑tower Sentence‑BERT (initialized with StructBERT) encodes queries and knowledge cards; contrastive learning (InfoNCE) aligns positive pairs while pushing apart negatives.

Two attention mechanisms (Query‑to‑Entity and Entity Self‑Attention) refine entity representations, and a fusion of original and similar query embeddings improves focus on central words. The final concatenated vector passes through dense layers for classification, outperforming standard BERT fine‑tuning.

Work Sharing – Case 2: Industry Search (采购商城)

Category prediction for product search is treated as a few‑shot text classification problem. Prompt‑learning converts classification into a masked language modeling task, enabling zero‑shot and ten‑shot performance with BERT‑base models. Self‑learning iteratively adds high‑confidence pseudo‑labels to enlarge the training set, achieving up to 82% accuracy on the full dataset and >90% in production after post‑processing.

Summary and Reflections

Key challenges include insufficient domain knowledge in short queries and scarce labeled data for specialized enterprise domains. Solutions involve internal knowledge‑card augmentation, few‑shot prompt‑learning, and potentially training enterprise‑specific large language models on internal data (e.g., ATA articles, contracts, code).

Future work also considers ensuring factual correctness of generative models by incorporating reinforcement‑learning style feedback and knowledge‑graph grounding before answer generation.