Advances in Knowledge Graph Construction and Applications at Alibaba's AliMe

Introduction

The AliMe team from Alibaba DAMO Academy shares recent advances in knowledge graph research, focusing on both domain‑specific and multi‑modal graphs and their deployment in e‑commerce scenarios.

01 Knowledge Graph Overview

Knowledge graphs model real‑world entities and relationships, typically using triples. They power search, recommendation, intelligent Q&A, and decision‑making. Applications are classified into two dimensions: the type of usage (raw graph retrieval vs. algorithm‑enabled) and the application form (business application, knowledge middle‑platform, or solution).

02 Domain Knowledge Graph Construction and Application

In dialogue‑driven e‑commerce, the goal is to connect users and products. Three core problems are identified: inferring user intent from questions, answering product‑specific queries, and providing explainable recommendation reasons. The team builds a POI‑centric graph linking user pain points, product attributes, and interests, using phrase mining, entity recognition, and relation extraction pipelines (including wide & deep features, BERT‑MLM filtering, CNN‑CRF, BERT‑BiLSTM‑CRF, and K‑BERT). The resulting graph spans 50 industries with millions of schema triples and billions of instance triples.

Applications include: (1) improving coarse‑recall in product recommendation by rewriting user queries into POI terms; (2) generating explainable selling points via graph aggregation and attention‑based decoding; (3) supporting live‑streaming scenarios such as virtual anchor script generation and short‑video creation.

03 Multi‑modal Knowledge Graph Construction and Application

To support live‑streaming, the team extends the domain graph with image and video modalities, creating the AliMe MKG. Key techniques involve text knowledge mining, image processing, and multi‑modal fusion using a pixel‑BERT‑style model that aligns OCR‑derived text tokens with CNN‑extracted image features. This enables text‑image matching without extensive annotation.

AliMe MKG now contains millions of textual nodes and tens of millions of image nodes for categories such as beauty, apparel, and retail. It powers virtual‑anchor script generation, short‑video production, and intelligent assistant features like product search, clarification, and multi‑modal Q&A.

04 Takeaways

Knowledge graphs provide "snow‑in‑the‑fire" value for cold‑start recommendation, content creation, and long‑tail scenarios, while also offering "icing‑on‑the‑cake" improvements when combined with existing models. The team emphasizes the importance of aligning graph construction with concrete business problems and differentiating between raw‑graph and algorithm‑enabled usage.

References

Key publications include AliMe MKG (CIKM 2021), AliMe Avatar (SIGIR 2021), AliMe KG (CIKM 2020), and several works on phrase mining, NER, and multi‑modal transformers.