Applying Knowledge Graphs to E‑commerce AIGC: From Domain‑Specific to General Knowledge Graphs and LLM Integration

The presentation introduces JD.com's exploration of AIGC in e‑commerce, outlining a multi‑modal input pipeline that combines product images, text, titles, and both domain‑specific and general knowledge graphs to generate controllable marketing copy.

Model architecture includes separate text and image encoders, a decoder employing restricted and copy mechanisms, sentence‑level fluency models, and punctuation correction to ensure coherent, high‑quality output across various lengths.

Four practical scenarios are demonstrated: product discovery channels, guide‑robot recommendations, social‑app sharing, live‑streaming scripts, and bundle‑purchase descriptions, all relying on accurate, multi‑modal text generation.

Controllable generation is achieved through three angles: input sanitization, token‑level vocabulary control, and model‑level adjustments such as encoder/decoder initialization and multi‑task learning, with a focus on preventing factual errors like incorrect attribute values.

Domain‑specific knowledge graphs are used to enforce attribute fidelity, improve copy mechanisms, and supplement sparse graph data via multimodal cues; visual gates (local and global) help the model attend to relevant image regions.

Large‑model research explores integrating domain and general knowledge graphs into LLMs, defining four knowledge categories and five pre‑training tasks to enhance NLU and NLG performance, resulting in higher ROUGE scores, longer, more diverse copy, and improved audit pass rates.

Experimental evaluations on e‑commerce tasks (knowledge‑graph completion, multi‑turn dialogue, product summarization) show that models enriched with domain knowledge (K‑PLUG) outperform generic baselines, confirming the value of knowledge‑graph‑augmented LLMs for real‑world e‑commerce applications.