Advances in Knowledge Graph Construction and Applications by Alibaba's AliMe Team

The article, prepared by Alibaba's DAMO Academy AliMe team, introduces recent advances in knowledge graph research, focusing on both domain-specific and multimodal knowledge graphs and their applications in e‑commerce.

Knowledge Graph Overview: Knowledge graphs model real‑world entities and relationships, typically using triples, and are widely used in search, recommendation, intelligent Q&A, and decision‑making.

Application Types: Two dimensions are discussed: (1) application categories – raw graph usage (search, recommendation, Q&A) versus algorithm‑enabled downstream tasks (decision analysis, creative generation); (2) application forms – business applications, knowledge middle‑platforms, and solution services.

Domain Knowledge Graph Construction: The team builds a domain KG to connect user intents (POI) with products, addressing three dialogue challenges: inferring user intent, answering product queries, and providing explainable recommendation reasons. Techniques include phrase mining (wide & deep features, Viterbi decoding, BERT filtering), entity recognition (remote supervision, lexicon augmentation, CNN‑CRF, BERT‑BiLSTM‑CRF), and relation extraction (BILSTM + GCN + CNN, BERT + FC + softmax, K‑BERT).

Multimodal Knowledge Graph (MKG): Extends the domain KG with image and video modalities to support live‑streaming e‑commerce scenarios such as virtual anchors and intelligent assistants. Construction involves text knowledge mining, image processing, and multimodal fusion using pixel‑level BERT models for image‑text matching.

Applications: • Dialogue recommendation – improves coarse recall by rewriting user queries into POI‑based queries. • Explainable recommendation – generates POI‑driven selling points using graph aggregation and attention‑based decoders. • Virtual anchor – automatic generation of text‑image scripts and short videos. • Intelligent assistant – multimodal product search, clarification, display, and Q&A using MKG.

Takeaways: Knowledge graphs provide essential support for dialogue and live‑streaming e‑commerce, enabling snow‑in‑the‑fire solutions (cold‑start, long‑tail) and value‑adding enhancements (knowledge infusion). Effective use requires clear business value assessment, appropriate application type (raw vs. algorithm‑enabled), and suitable deployment form (business, platform, solution).

References: The article cites several Alibaba publications on AliMe KG, MKG, KBQA, and related AI techniques, as well as external works on phrase mining, NER, K‑BERT, and multimodal transformers.