Recently, Li Biao, the Algorithm Center of Tencent PCG News Product Technology Department, was invited to Tencent Media Research Institute for internal sharing, and sorted out the algorithm application product scenarios in detail. The following is a partial record. The topic I share with you today is algorithm-enabled content processing and distribution, focusing on content processing. Before we start, let’s introduce the application scenarios of the algorithm in Tencent News. The first is the understanding and distribution of various content forms (such as text, video, audio, topics, Q&A, etc.) in Tencent News APP, which involves recommendation systems and AI algorithms to enable content operations. The second is to push Tencent.
News to WeChat, each time a big phone number list picture and three news information, a total of four, click on some bottom pages to jump to the Tencent News APP. The third, Dolphin Smart Sound, a "listening to the news" artifact, mainly used in smart speakers, car audio and smart home appliances, can currently provide 70% of the voice information on the market; it involves voice summaries, voice recording and personalized voice recommendations algorithm. The fourth, Dreamwriter, involves many things such as manuscript writing, content creation, draft screening, and matching pictures, which is also the focus of this article. First, the framework of the algorithm The overall framework of the algorithm consists of the underlying algorithm and the upper-level application. The underlying algorithms include NLP lexical, syntax, text understanding, etc., visual image quality, image-text matching.
There are two types of applications grafted on top of the underlying algorithm, namely recommendation system and search. The recommendation system can be divided into five steps. 1. Content Processing It is also called a content management system, which embeds algorithms related to text classification, tagging, abstract extraction, semantic analysis, content de-duplication, content analysis, error correction, image matching, draft screening, etc. and content processing. 2. Index The content after the primary selection, that is, information such as pictures, texts, videos and other information that is ready to be distributed to users for consumption, is added to the index. 3. Portrait It can be divided into two parts: basic portrait and extended portrait. Basic portraits provide users with personalized recommendations based on user classification, points of interest such as tags, user basic attributes.