Dr. Zhou Ming is the vice president of Microsoft Research Asia, the president of the International Association of Computational Linguistics (ACL), the director of the Chinese Computer Society, the director of the Chinese Information Technology Committee, the former director of the terminology working committee, and the executive director of the Chinese Information Society. Harbin Institute of Technology , Tianjin University, Nankai University, University of Science and Technology, Beihang University and many other doctoral tutors.

The following is a dialogue with Dr. Zhou Ming. The heart of the machine has been streamlined and edited according to the content of the dialogue without changing the original intention.

Heart of the machine: According to Marekrei, in the past year (2018 years), the number of papers you participated in published wasNLP+ML field ranked first in the world. Can you tell us what achievements your team has achieved in the past year or so?

Zhou Ming:Overall, last year was a big harvest for our team. Our achievements can be summarized as follows:

The first achievement is reading comprehension (MRC). We have achieved the first place in both SQuAD 1.1 and SQuAD 2.0.2018 1 Month, our submitted system broke through the human level at the SQuAD 1.1 test set for the first timeLater, several other companies have surpassed human scores. These systems reflect the breakthrough in reading comprehension. Recently, we surpassed other companies on SQuAD 2.0 and won the first place. In addition, on an interactive, multi-round MRC system CoQA, we also received the first result (and also the first submission).

The second achievement is Neural Network Machine Translation (NMT). weOn the Chinese-English test set of the general news report test set newstest2017, it reached a level comparable to human translation.. This is the first translation system that can be translated manually in the quality and accuracy of news reports. This is a result of a collaborative effort between colleagues from the Natural Language Computing Group and the Machine Learning Group at Microsoft Research Asia and the machine translation product division at Microsoft headquarters. We propose new joint training and dual learning to make full use of monolingual corpus, and then invented consistency specifications and improved network decoding capabilities. The combined application of these technologies has greatly improved the level of translation.

The third achievement is the Grammar check. We use the neural network coding and decoding technology, and adopt a technique similar to neural network machine translation. It has made important improvements to the grammar check, and can automatically generate training corpus and decode it one by one. Our results ranked first in the three public review sets of current grammar checks. We publishedRelated ACL articlesCaused the attention of the industry.

The fourth is neural network based speech synthesis (TTS). We worked with Microsoft's Voice Products Division to apply neural network machine translation technology to TTS for the first time, greatly improving the quality of TTS. Our technology performs best in the relevant evaluation collections.

In addition, we continue to cooperate with Microsoft Xiao Bing. Based on the original chat engine, the machine creation ability has been greatly improved. For example, writing poems, composing music, news, etc., in which AI composing music is also on the CCTV's "tactful" program, which has a great influence.

Heart of the machine: In the past few years, you have been actively involved in the organization and management of the NLP Top ACL, especially this year as the chairman of the ACL. In the process, first of all, what do you think is the new development in the NLP field in the past year?

Zhou Ming:First, the neural network has penetrated into various fields of NLP, and the new methods of modeling, learning, and reasoning of neural NLP have made good progress in the typical NLP tasks I have just introduced. Second, toBERTA series of pre-training models represented by them have been widely used, reflecting the universal language laws and the potential combination of knowledge and specific application scenarios contained in large-scale linguistic data. Third, the low-resource NLP tasks have received extensive attention and A very good development.

In addition to the above-mentioned remarkable advances in technology, I think it is worth mentioning that the rapid progress of China's NLP has attracted worldwide attention. Major societies such as the Chinese Computer Society and the Chinese Information Society have made important contributions to the development of NLP in China. Each of the two institutes has a good academic conference, workshop or summer school. In addition, the two institutes also co-organized the "Language Intelligence Summit", which was the third year last year. Due to the efforts of these societies, coupled with the efforts of schools and companies, China's natural language development, from the publication of the top meeting (ACL, EMNLP, COLING, etc.), has ranked second in the world for the past five years; second only to The United States is much higher than other countries.

One more thing to say here is that the International Society for Natural Language Processing and Chinese Computing (NLPCC) of the Chinese Computer Society is gradually becoming closer to the world's top conferences in terms of its degree of internationalization, scale and level. The admission rate is around 23%, and the number of participants is above 500 and there is an increase of 20% year by year. At the same time, the conference has an international conference organizing committee and program committee, and the working language is English. It can be expected that it will become China's leading international NLP academic conference. At the NLPCC conference, we have stated two goals: 2020, China's natural language research reached the world's recognized advanced level; in 2030, China's natural language research reached the world's recognized top level. I believe that the NLPCC conference can promote the realization of these two goals.

In addition, last year ACL also established the Asian ACL Branch (AACL). I am very grateful to the ACL Executive Committee for its support and the support of NLP colleagues in various countries and regions in the Asia Pacific region. The establishment of AACL marks that Asia can make progress in NLP development in North America and Europe. After the establishment of the AACL Asia Chapter, many activities like ACL can be organized in Asia to improve the level of NLP development in Asia.

So you see, from China to Asia to the world, the whole trend is to work at different levels; China's progress is very rapid, attracting the attention of some researchers in the world. I pointed out in an article I wrote earlier.NLP entered the decade of gold. This is because of the huge demand for NLP brought by the development of national economy and artificial intelligence in the future. Large-scale data can be used for model training. Various new methods represented by neural network NLP will gradually improve the modeling level. A variety of evaluations and the ability of various open platforms to promote NLP research and promotion, and the growing prosperity of AI and NLP to promote the development of specialized talents. Therefore, the next decade is very worth looking forward to.

Heart of the machine: Just now you mentioned that China's NLP is progressing rapidly. In the past year, what are the more important breakthroughs in China's research?

Zhou Ming:In addition to the results of the Microsoft Research Asia mentioned above, many schools and companies in China have made good progress, such as Chinese MRC, Baidu, and Keda Xunfei have organized respectively with relevant institutes or universities. Large-scale Chinese MRC review. Its influence also transcends national borders. Other countries only need to do Chinese MRC, but also to participate in these evaluations.

Chinese machine translation, which is a Chinese-centered machine translation, is now at the leading level in the world. Chinese is the center, which is the translation of Chinese to other languages, Chinese to Japanese, Thai, Malay, etc. These are all centered on Chinese. There are many schools and companies in China that have done a terrific job around the translation of languages ​​in various countries along the Belt and Road. Good progress has been made in the research and practicalization of Chinese-centered research.

In terms of chat and conversation. China is also in the forefront of the world. The chat system represented by Microsoft Xiao Bing in China has led to the research and development of artificial intelligence chat systems worldwide. Xiaoyan’s average number of rounds of chat reached the 23 round, and multi-modal chat was achieved. In addition to Xiao Bing, many Internet, e-commerce, and mobile phone companies in China have developed research on chat bots and voice dialogue systems. Used in search engines, voice assistants, smart speakers, Internet of Things, e-commerce, smart homes, etc.

These three trends, I think, represent some typical advances in China's NLP over the past two or three years.

Heart of the machine: In the future, what kind of research do you think will have greater research potential in 2019?

Zhou Ming:If you want to count, I think there are three points that are more interesting.

First, the pre-training model just mentioned. Basically everyone has been talking about pre-training models in the past year. In particular, when BERT came out, almost all tasks used BERT, and as a result, the level of many tasks was improved. So what I can expect in the coming year is that the pre-training model will continue to heat up. This includes how to train a better pre-training model, including how to better apply the pre-training model to a specific task.

Second, the study of low-resource NLP tasks. How to do certain learning, modeling and reasoning in the absence of corpus or small corpus? Further develop semi-supervised learning, unsupervised learning methods, and use Transfer Learning, Multi-task Learning, etc. to cleverly graft or borrow models from other languages, tasks, or open fields into new languages, tasks, or fields. A specific task (such as machine translation, reading comprehension, question and answer, etc.) is better reflected.

Third, it is called application based on knowledge and even based on common sense. It is how to build knowledge and common sense, how to subtly integrate into the model, and then how to evaluate the effects of knowledge and common sense. I think this may become a concern in the coming year.

Heart of the machine: You pointed out in the article "NLP will usher in the Golden Decade" just mentioned that NLP will tilt to four aspects: 1) to introduce knowledge and common sense into current data-based learning systems. Medium; 2) low-resource NLP task learning method; 3) context modeling, multiple rounds of semantic understanding; 4) based on semantic analysis, knowledge and common sense interpretable NLP. These are just mentioned by you. What are your research ideas in these areas?

Zhou Ming:First of all, in terms of methodology, the pre-training model I just mentioned, as well as semi-supervised learning, transfer learning, and multi-task learning, we will do our best to advance. Then in the specific application, we will take machine translation, reading comprehension, question and answer, chat conversation, and possibly some other aspects as a starting point. Then for these specific tasks, implement the methods just mentioned, look at the effects, and then iterate.

You see that we have two legs to walk, one is the method, the other is the application, let it keep iterating. Applications provide challenges for many methods, and then many methods provide new ideas for the application, and the two can complement each other.

In addition, on the one hand, we expect the project to do well, the application itself is well done, can serve a lot of users, through Microsoft products, or dedication to the open source community. Use user feedback to continuously adjust and improve our research direction and ideas. On the other hand, methodologically, we hope to clearly describe the theoretical system of some tasks (such as natural language understanding) under the new neural network architecture, including modeling, learning, and reasoning. The three ing things are actually natural language as a discipline, establishing the most important technical system and theoretical system behind it. Now there are good studies in these three aspects, but scattered in many places, not a complete system, so we hope to answer natural language (especially based on neural network computing) through project and research. What is the theoretical system of language). Under the support of the theoretical system, I hope to finally form a technical system. We will release tools or open source systems to help people working in natural language research at home and abroad, so that they can quickly learn from existing methods, do not repeat research, and use time to do some applications that are of interest to them.

Heart of the machine: Your team has also extensive research on multimodal fusion and has published several papers. What progress has this research field made now?

Zhou Ming:First of all, multimodal fusion is very interesting. Due to the progress of the neural network, the encoding and decoding of multi-modality (language, text, image, video) can be unified under the same framework. Because the intrinsic semantics are different, it is really not clear how the results of the language analysis can be combined and applied together with the results of the image analysis; now a model can be used to model, encode, and decode. Thus, end-to-end learning can be integrated and unimpeded.

Second, the application, and correspondingly produced a lot of interesting applications, such as capturing, is to understand a picture or video and then use a paragraph of text to describe it. There have been many such studies in the past year or two. There is also a question or answer about video or images, (CQA). CQA In the past 1 to 2 years, there has been a lot of progress, including our group also doing some CQA work, such as introducing common sense knowledge to help improve the level of CQA.

The third is to use the result of image recognition as the input of the natural language system to do the work of writing poetry, lyrics and composing music. Microsoft Xiaobing wrote poetry as well. The user uploads a picture, Xiao Bing understands the picture, and the results of the understanding can be represented by several keywords. Then use keywords to generate more associative keywords, and then generate a lyric or a poem.

The heart of the machine: A few days ago, I read an article by Teacher Feng Zhiwei entitled "Languageists have a lot to do in the study of natural language processing." What role does linguistics play in the history of NLP? At present, does linguistics still have an effect on NLP?

Zhou Ming:The article written by Teacher Feng is to look at this issue from a linguistic point of view. I think the angle is very good and we are also concerned about these issues.

In the past, natural language processing was based on linguistics at the beginning, so linguistics played an important role in the development of natural language processing, including the important contribution of rule-based NLP systems to syntactic analysis and machine translation.

But in the past few years, we have also noticed that because big data (labeled data) is getting easier to get, end-to-end training can be done through machine learning (statistical machine learning or neural network learning). If you only look at the results, if you have enough data, it seems that you can get good results without the knowledge of linguistics. For example, neural network-based machine translation does not use linguistic knowledge. This is a trend that everyone is currently seeing.

But this does not mean that linguistics is really useless. For example, the translation of low-resources, at this time bilingual corpus is very small, the machine translation system obtained by conventional machine learning, its translation quality and generalization ability are very insufficient. At this time, you can consider incorporating linguistic knowledge into it and hope to get a better translation result. In this sense, linguistic knowledge plus human domain knowledge can certainly play a role in semi-supervised learning or low-resource natural language tasks.

However, I don't think there is a particularly good way to combine the two, nor to fully express linguistic knowledge or domain knowledge to reflect its capabilities. There is still a lack of research in this area, but it is also a research focus in the future.

Is linguistics useful for NLP research? I feel that I need to look at this situation by case by case. It cannot be simply said to be useful or useless. It is necessary to consider various methods in terms of the number and quality of the specific tasks and resources (data, knowledge and rules) that can be obtained. For example, if the data is sufficient enough to be end-to-end automatic learning, the data is not sufficient, and the available knowledge and rules are available, there is no reason not to use knowledge and rules to quickly build the system. When the system is up and running, you need to consider constantly adding data, knowledge, and user feedback to improve the system. So a practical NLP system is a combination of data, knowledge and users.

机器之心：我们注意到微软亚洲研究院从1998年到去年正好20年；而您是1999年加入微软，到2019年也正好20年。这期间有哪些让您觉得特别难忘的事情？

Zhou Ming:Last year, Microsoft Research Asia 20 anniversary, we held a series of events, invited a lot of old friends (including academia, industry) and colleagues from Microsoft headquarters, everyone gathered together to talk. This time is just a time to look back and look to the future. So everyone thinks about what happened in the past 20 years, and what left us a deep impression. I also thought about it here.

I think that in the past, 20 was first and foremost a growing 20 year. The entire Microsoft Asia Research Institute went from scratch, from small to large, and experienced a turbulent process. Sometimes we develop very smoothly, and sometimes we experience some setbacks. But no matter what, we are always making progress and getting better and better.

Secondly, I am honored to have gone through all the processes led by the first Dean, Kai-Fu Lee, and now President Hong. So for me, I am both a witness and witness of this 20 year, and more meaningfully, I am a beneficiary and learner. I learned a lot of things in 20. I joined Tsinghua from Microsoft and found that there are really a lot of fresh things to learn. Microsoft has a strong product and marketing team, as well as a strong research atmosphere at Microsoft Research and Microsoft Research Asia. As an employee, I have gained a good opportunity to learn and experience from all angles. Specifically, regardless of research methods, teamwork, product awareness, and the realm of domestic and foreign cooperation. I feel that I have been well experienced in these areas.

If you are talking about a project, let's talk about our history with a few examples. We started with Microsoft's input method, Chinese and Japanese. In the 2004 year, I started to work as a Microsoft couplet. (Note: The NLP team of Microsoft Research Asia belongs to the earliest do couplet and poetry in China). From 2008 to 2012, we did the Bing Dictionary, and the founder of Microsoft Research Institute of 2012. Rick Rashid demonstrated the real-time voice machine translation system at the 21st Century Computing Conference. In the last two or three years, we participated in the cooperation of Microsoft Xiaobing. In the past few years, we have also done neural network machine translation, as well as machine reading comprehension and so on. I think every project has its own characteristics. From the beginning of the project, through the constant program adjustment, experiment with different methods. In this process, everyone has gained a good experience and improved their research capabilities. Looking back at these things and going through one project after another, my colleagues and I are very proud and proud.

However, I want to project the perspective from the pure research project itself to a broader world related to the development of NLP. In fact, the past 20 years of Microsoft Research Asia played a unique role in promoting global NLP, especially China's NLP. As a big company, a responsible company should not only think of itself, but also think about whether it can positively help the development of this field and help its countries and regions develop in this field. Be a meaningful contributor. In the field of NLP, when Microsoft China Research Institute (later renamed Microsoft Asia Research Institute) was founded, there was only one ACL article in China, written by the research team of Tsinghua University's Huang Changning.

At the beginning of the establishment of Microsoft Research Asia in 1998, we have developed a plan for cooperation with relevant schools and schools, and jointly promoted the research level of NLP through summer school, joint laboratories, academic conferences, and various university cooperation projects. In the past 20 years, we have trained more than 500 interns, 20 doctoral students, and 20 postdoctoral fellows in the NLP field. Most of these people have gone to school or other companies. These people have gradually become the technical core or leader of their unit; they have also led to the growth of more people. So after 20 years, China's NLP is getting better and better, and it has ranked second in the world in NLP summits (such as ACL) for five consecutive years.

Now the NLP group of major Chinese companies is led by world-class experts and serves world-class tasks. Whether it is an article or a product, it is a world-class level. Behind this fact, I would like to say that there is a huge relationship with Microsoft Asia Research Institute in China that has promoted the development of this field at the right time.

Microsoft Research Asia is known as the Whampoa Military Academy in the IT or artificial intelligence world; we can also proudly say that we are also the Whampoa Military Academy in the NLP field. So every time I read this, I feel very proud and proud.

This article is transferred from the public magazine Microsoft Research Institute AI headlines,Original address