On the evening of June 3^rd, 2019, the Institute of International and Strategic Studies (IISS), Peking University (PKU) held the 37^th "North Pavilion Seminar". Zhao Xiaobing, Professor of the School of Information Engineering, Minzu University of China and Permanent Member of the Chinese Language Modernization Society, gave a seminar entitled "Language and Intelligence", expounding the history, current situation, and future of natural language processing in simple terms. The seminar was hosted by Gui Yongtao, Vice President of the IISS and Vice President of the School of International Studies (SIS), PKU.

The seminar elaborated the topic from four aspects: processing of language and natural language, artificial intelligence and deep learning, challenges and prospects of artificial intelligence, and the work being carried out by her institute.

First, Prof. Zhao Xiaobing started the topic with the introduction to "processing of language and natural language", explaining how natural language processing serves as a bridge between human natural languages and computer machine languages and achieves the interconnection between them through comprehension, conversion, and generation. Taking machine translation as an example, Prof. Zhao introduced three stages of its development: (1) the development based on grammatical rules in the 1980s, (2) the development based on the statistical method (statistical machine translation, SMT) in the 1990s, and (3) the development based on neural networks (neural machine translation, NMT) after 2010. The mode based on neural networks is currently the mainstream approach to machine translation. Benefiting from the dramatic enhancement in computer performance, the mode is freed from the traditional mode of "input-segmentation-translation-reordering", and is committed to establishing mapping relations between source language and target language directly by using a large corpus of millions or even tens of millions of bilingual sentence pairs. The method has now gone out of the laboratory, and its results have been accepted by the market.

Secondly, Prof. Zhao analyzed the current achievements and limitations of artificial intelligence from the perspective of "deep learning". There are three waves during the development of artificial intelligence: (1) the stage of rule-based reasoning in the mid-20th century, (2) the stage of machine learning in 1980s, and (3) the stage of deep learning at the beginning of the 21st century. At present, artificial intelligence has made breakthroughs in the field of "perception", but it still struggles in the field of "cognition". The essence of deep learning is "induction". Based on the neural network structure of multiple nonlinear transformation, machines can abstract and learn identifiers of data from phenomena (such as text, speech, and images), and they do even better than humans in well-defined fields. But they cannot perform deductive reasoning, and cannot understand things behind language like humans do. For example, machines can extract the vector features of "cat" from numerous pictures and identify the "cat", but if you input "an animal with four legs and whiskers", machines cannot construct the "cat" image. Another example is that machines can answer who the U.S. president is, but cannot answer why the U.S. president could not sleep last night. Therefore, deep learning is not the only thing artificial intelligence can do, but a means for artificial intelligence to realize human intelligence. Using machines to simulate human learning behaviors still has a long way to go.  (Contributed by Dong Rong) 

Editor: Li Fangqi, photography: Zheng Peijie