Intelligent Online Selling Point Extraction for E‑Commerce Recommendation (IOSPE) Wins AAAI 2022 Innovation Award

At the 36th AAAI 2022 conference, JD.com’s Retail Search and Recommendation platform won the AAAI Artificial Intelligence Innovation Application Award for its project “Intelligent Online Selling Point Extraction for E‑Commerce Recommendation” (IOSPE). The award recognized the system as the sole Chinese team to receive two AI innovation awards at the event.

IOSPE is an intelligent selling‑point extraction system that leverages natural language processing and deep generative techniques to automatically produce recommendation copy for products. By September 2021, IOSPE had become a core service for over 60 key product categories, covering more than 4 million items and generating over 110 million selling points, serving multiple JD.com platforms such as the main site, JD Joy, the fast‑track version, and promotional pages.

To date, IOSPE has generated over a hundred million selling points, increasing product click‑through rates (CTR) by more than 2 % and raising average user dwell time on product pages by over 0.32 %, a significant boost for large‑scale e‑commerce.

The core AI technologies consist of two main stages: (1) extraction and generation of short selling‑point texts using product details and user reviews, and (2) personalized distribution of the generated points based on user profiles. The extraction pipeline includes three steps: coarse filtering with a BERT‑based scoring model, generation with a transformer‑pointer network, and fine filtering with a refinement model.

Figure 1: Multi‑angle selling‑point themes help cover a wide range of products.

Figure 2: Recommended selling points empower multiple application scenarios.

The overall IOSPE workflow (Figure 3) begins with gathering source material—product attribute tables, titles, detail‑page text, and positive user reviews—feeding them into the extraction and generation module to produce several high‑quality selling points per product (e.g., a pomegranate). These points are then personalized and distributed to different customers based on their interests.

Figure 3: IOSPE overall process diagram.

The extraction and generation pipeline (Figure 4) comprises: (a) a BERT‑based coarse‑filtering scoring model that ranks raw text snippets, (b) a transformer‑pointer network that generates selling‑point sentences from the top‑ranked snippets, and (c) a fine‑filtering model that selects the highest‑quality outputs.

Figure 4: (a) Coarse‑filter model, (b) Generation model, (c) Fine‑filter model.

The system is deployed within JD.com’s recommendation architecture (Figure 5). When a request arrives, the SOA platform triggers the front‑end (Broadway) to collect user profile, purchase history, and product information (attributes, reviews, descriptions, images). This data is sent to an indexing module, which forwards it to the recommendation engine. The AI‑flow component performs recall and ranking, filters candidates by inventory and popularity, and finally passes the selected products to the selling‑point module for extraction and personalized distribution.

Figure 5: Overall architecture deployment of the recommendation and selling‑point system.