Constructing Knowledge Graph for Social Networks in A Deep and Holistic Way

Abstract

Online social networks such as Facebook and LinkedIn have been an integrated part of everyday life. To improve the user experience and power the products around the social network, Knowledge Graphs (KG) are used as a standard way to extract and organize the knowledge in social networks. This tutorial focuses on how to build KGs for social networks by developing deep NLP models, and holistic optimization of KGs and the social network. Building KG for social networks poses two challenges: 1) input data for each member in the social network is noisy, implicit and in multilingual, so a deep understanding of the input data is needed; 2) KG and the social network influence each other via multiple organic feedback loops, so a holistic view on both networks is needed. We will share the lessons we learned from tackling the above challenges in the past seven years on building the Knowledge Graph for the LinkedIn social network. To address the first challenge of noisy and implicit input data, we present how to train high precision language understanding models by adding small clean data to the noisy data. By doing so, we enhance the-state-of-the-art NLP models such as BERT for building KG. To address multilingual aspect of the input data, we explain how to expand a single-language KG to multilingual KGs by applying transfer learning. For the second challenge of modeling interactions between social network and KG, we launch new products to get explicit feedback on KG from users, and refine KG by learning deep embeddings from the social network. Lastly, we present how we use our KG to empower more than 20+ products at LinkedIn with high business impacts.