321, the JD Yanxi × NLPCC 2022 Challenge begins!

With the rapid development of the Internet, the problem of information overload has been brought about, which has brought obstacles to people's acquisition and understanding of the information they need. As a result, there are many essential and summary service models, such as Luo Zhenyu's 60-second voice, quick book, quick movie, etc., all of which serve the audience by extracting core information. In recent years, AI technology has made great progress, especially in the field of text generation in natural language processing. It has a very wide range of industrial application prospects. It can be based not only on text, but also on the basis of multi-modal information such as pictures to condense and generate beautiful copywriting. . The 175 billion-parameter GPT-3 released by OpenAI in 2020 has a writing level comparable to that of humans.

JD.com finds good products-Siemens copywriting

JD.com found a good product - sea blue mystery copywriting

JD.com has applied text generation technology on a large scale in its own business. It can write exquisite product summary texts for users in a short time, and recommend it to users to "plant grass" according to user preferences. JD.com pre-trains K-PLUG based on the domain The product copy generation model has covered more than 3,000 third-level categories of JD. The abstract copy reflects the most unique and differentiated value of the product, maximizes the promotion of user purchases, and brings a cumulative GMV of more than 300 million yuan. In addition, Yanxi can also have various AI capabilities such as writing poems, writing Spring Festival couplets, and writing calligraphy.

Jingdong found good goods - horizontal refrigerator copywriting

The challenges of product abstract copywriting mainly come from three aspects. First, the sources of information are rich, including titles, product specifications, product posters and other information; second, multi-modal and structured information needs to be processed, including text and images. , and structured information such as product specifications and parameters; thirdly, the AI system must be intelligent enough to fully mine the selling point information of the product, and recommend it to the right user at the right time.

Based on this, JD Yanxi has opened up the desensitization data of JD.com's real scenes, and joined hands with NLPCC 2022 to hold a multi-modal commodity abstract challenge. The task is to generate a condensed text summary for a specified product. The given input includes product detailed text description, product knowledge graph and product image. The definition, dataset, and evaluation method for this task are described below.