NLG Solutions for E‑commerce: DAMO Academy’s XGeneration, Word2Text, KG2Text, and Metaphor Generation

Background – E‑commerce platforms demand large‑scale, high‑quality promotional content, but merchants often cannot keep up with the rapid pace of platform activities, creating a need for automated or semi‑automated content generation.

XGeneration – A content‑reproduction solution offering high automation and low merchant effort, adaptable to various e‑commerce scenarios.

E‑commerce NLG Practice

Controllable short‑text generation (RQ1: attribute control, RQ2: logical control) using token‑to‑text methods.

Word2Text models: PKM (baseline BART, dataset of 300 k industry‑labeled copy) and PSCN (enhanced with twin encoders, dataset of lexical‑controlled product descriptions).

KG2Text model: builds a knowledge‑graph with product attributes, demand nodes, and user attributes to ensure logical consistency.

Metaphor Generation: produces metaphorical sentences by first generating a base metaphor and then a supplementary clause, improving appeal.

New Media Exploration

Short‑video script generation – automatically creates scene numbers, shot types, durations, visual descriptions, dialogues, and sound effects, reducing script‑writing time by up to 60 % and achieving high adoption rates.

Intelligent short‑video editing – combines product metadata, selling points, multi‑modal retrieval, and rendering pipelines to produce complete videos with strong user engagement.

Results – The models achieve >80 % rationality and fluency, 64 % script adoption, 98.6 % public‑video pass rate, and significant improvements in playback and conversion metrics during major sales events.

Conclusion – Two complementary solutions (NLG practice and new‑media exploration) address the e‑commerce content‑reproduction problem, enabling scalable, high‑quality content generation without manual effort.