Research Center for Cognitive Computing and Natural Language

The Research Center for Cognitive Computing and Natural Language (CCNL) is committed to promoting further development of the new generation of fundamental cutting-edge technologies of cognitive computing and natural language represented by large pre-training models. The research center endeavors to solve all technical problems during the commercialization of large model technologies, to build new technical structures for natural language fields such as dialogue robots, knowledge extraction and knowledge systems, and to develop new technological paradigms for cognitive AI.

Dr. Zhang Jiaxing serves as Chair Scientist at the Research Center for Cognitive Computing and Natural Language. He obtained his PhD from Peking University and served as Researcher in Microsoft Research Asia, Senior Expert of Algorithms in Ant Financial and Chief Scientist in 360 DigiTech. Dr. Zhang Jiaxing has published more than 20 academic papers at top-level conferences and journals (NIPS, OSDI, CVPR, SIGMOD, NSDI, AAAI, WWW, etc.) of a number of fields, including natural language processing, deep learning, distributed systems and physics, and filed more than 70 patents. He set up the deep learning team during his service in Ant Financial, taking the lead to apply deep learning in industrial scenarios. He also developed the AI directions of natural languages represented by Ant Financial’s intelligent customer services, intelligent investment consultancy, intelligent post-loan services and intelligent public opinions. In addition, he developed a centered platform for the integration of data AI in 360 DigiTech, which has become a fundamental framework of fintech.

Cognitive abilities of machines are to make machines listen, speak, read, write, learn, think and understand the world like humans. This is the foundation of the AI industry as well as the high ground of AI technology. A machine equipped with cognitive abilities is able to replace human brainwork in many scenarios, such as customer service, consultancy, marketing, medical diagnosis, education, financial investment, caregiving and companionship. In these areas, machines can free humans from labors or enhance the efficiency of humans and inject massive “cognitive labor forces” into society. Apart from that, cognitive abilities of machines point to the ability of semantic understanding at a deeper level and can improve the level of search, recommendation, advertising and big data technologies fundamentally, thus triggering a new round of algorithmic revolution. Lastly, the enabling of machine to “cognize” the world means that algorithms can understand the world as humans do, possibly at a more comprehensive and deeper level. Such cognitive abilities of machines provide assurance for better regulation of economy and governance of society and thus lead to a brighter future for mankind.

Research Findings

The Research Center for Cognitive Computing and Natural Language is developing the new generation of natural language cognition platform that is based on large pre-training models to empower the whole ecosystem of natural language technology with machine cognitive abilities brought by the large models and thus promoting technological fairness. 

The first feature that the cognitive platforms integrates is the large model training capability. Underlying that is a complete pipeline with constant production and update of large models to support the platform functions. The large models can be directly exported to the partners of the ecosystem. Built on those large models are learning engines for various natural language processing (NLP) tasks, including classification, sequence marking, sentence relationships and text generation. Users may learn product level models for specific tasks from a large model by providing few examples (few-shot learning) only. Behind that are various technologies such as fine-tuning, model distillation, model quantification, model structure pruning and searching, as well as deep optimization and binding of them with specific NLP tasks. In addition to the task learning for specific NLP tasks, the cognition platform provides a deep learning framework that is based on large models, allowing partners of the ecosystem to further develop models based on this framework. 

Research Directions

Cognitive Machine Learning Platform
Based on large pre-training models, this direction builds task learning engines for different NLP tasks, helping users to get models with better effects at lower training costs. 
Dialogue Robot Framework
Based on the latest natural language technology represented by large models, this direction combines semantic understanding engines, knowledge systems and simulated audio technology to build the next generation of dialogue robot framework and trigger the ecosystem of chatbot. 
Knowledge Extraction Engine
Based on semantic models and gigantic knowledge systems, this direction enables machines to extract information from sentences and articles and helps machines better understand the world. 
Knowledge System
Knowledge is the ultimate obstacle to overcome in cognition. Through systematic design, this direction builds knowledge into a large system that integrates different segments, including structured knowledge (knowledge graphs), knowledge acquisition, knowledge application and knowledge validation, which serves as the underlying infrastructure of cognitive intelligence.