Applying Knowledge Graphs to E‑commerce AIGC: From Domain to General KG and Large Language Models

The presentation introduces JD.com’s exploration of AIGC in e‑commerce, outlining a multi‑modal input pipeline that combines product images, textual details, and both domain‑specific and generic knowledge graphs, which are transformed into attribute‑value pairs for downstream generation.

After preprocessing, the data flow enters a model consisting of separate text and image encoders and a decoder that employs constrained and copy decoding, sentence‑level fluency models, and punctuation correction to ensure high‑quality, coherent product copy.

Four practical scenarios are demonstrated: the "Discover Good Products" channel, a guide‑robot for customer service, social‑shopping app content, and live‑streaming copy generation, illustrating the system’s ability to produce short, medium, and long‑form marketing texts.

The article then details techniques for controllable text generation, including input sanitization, token‑level biasing, and model‑side adjustments such as modified initialization and multi‑task learning, emphasizing the importance of fidelity, diversity, and readability in e‑commerce contexts.

Domain‑specific KG‑driven generation is explored through examples like washing‑machine specifications, showing how direct KG lookup can guarantee accurate attribute values, while a refined copy mechanism restricts generation to tokens present in the input KG.

Integration of generic KG enriches copy with broader knowledge (e.g., usage scenarios for instant noodles), and token‑type embeddings help the model selectively trust domain versus generic sources, improving relevance and reducing hallucinations.

Large‑model research is presented, describing the development of a knowledge‑enhanced LLM (K‑PLUG) that incorporates four knowledge types (product KG, element schema, category, and selling points) via specialized pre‑training tasks, leading to significant gains in NLU and NLG benchmarks such as ROUGE and human‑rated fidelity.

Experimental results demonstrate that knowledge‑infused models achieve higher fidelity (up to 93% correctness), longer and richer descriptions, and better audit pass rates compared to baseline T5 models, confirming the value of knowledge‑grounded pre‑training for commercial text generation.

The talk concludes with reflections on future directions for controllable generation, multimodal KG completion, and scaling LLMs with knowledge integration to further improve e‑commerce AIGC applications.