Editor's Note: "What was it like to study for a PhD at Microsoft Research Asia?" "Students with scientific research dreams gathered at Microsoft Research Asia from major universities to receive joint guidance from mentors from both universities and research institutes. Bottlenecks, setbacks and difficulties are often similar in doctoral quests, but everyone has their own experiences, experiences and answers to growing up and finding themselves.
"Teach yourself Python in one day, TensorFlow in three days, and reproduce the results of a top article in three days." Recalling his selection for the joint doctoral training program of Microsoft Research Asia and Sun Yat-sen University, Zhong Wanjun, who has been in the fourth year of his doctorate, still remembers it vividly.
Time flies, and the girl who tasted scientific research at the beginning of the year achieved an important scientific milestone this fall: being named the "Microsoft Scholar" of 2021. As a representative of "Microsoft Scholars" in Asia, she elaborated on her research vision in the Microsoft Global Winner Announcement Video.
MSR Asia Fellows - Wanjun Zhong - Fashion - HD Full Genuine Video Watch Online - Youku
From being rejected seven times to being named a "Microsoft Scholar", how did Wanjun Zhong, an intern in the Natural Language Computing Group of Microsoft Research Asia, continue to grow in the face of scientific challenges? From her story, I believe you will be able to draw nourishment about "growing up in the challenges of scientific research".
"Taking the initiative" and "challenging the impossible" in scientific research
"The research environment and research resources of Microsoft Research Asia are very attractive to me." Zhong Wanjun recalled his curiosity and excitement about the institute when he sat in the lecture hall of the joint training doctoral program in the summer of 2017, and the life trajectory that was originally planned to study abroad was also reversed to a new course because of this curious registration seminar.
According to the training plan, the selected joint training doctoral students will spend their senior year as interns at Microsoft Research Asia, return to campus for the first year of their doctorate to complete degree credits, and explore cutting-edge topics in Microsoft Research Asia for the next four years under the joint guidance of senior researchers and university tutors at Microsoft Research Asia.
Zhong Wanjun's doctoral first-year course only took half a year. "I especially wanted to come back early to do scientific research with everyone, so I ranked all the credits in the first half of the semester." In 2018, he joined the Natural Language Computing Group of Microsoft Research Asia and began to explore the application of machine reasoning technology to natural language tasks. Zhong Wanjun has a strong initiative at every step of the road of scientific research: she will take the initiative to fight for her favorite research direction and explore more suitable research topics for herself; when the scientific research status is not good, she will also take the initiative to analyze the reasons and then quickly seek solutions.
After three years of doctoral research, Zhong Wanjun has published several high-quality papers in top natural language conferences such as ACL and EMNLP, and more than half of the papers are the first authors, and these works have also had a certain impact in related fields.
These scientific results are not achieved overnight, and they are accompanied by challenges that seem "impossible". In 2020, she participated in a project aimed at complex reasoning problems in the Entrance Examination (LSAT) of american university judicial colleges, which brought her the most scientific research experience and self-breakthrough.
"It's really, really difficult." She recalled. The "currently no good solution" problem is to use machine learning methods to solve the most difficult part of the U.S. Judicial College entrance exam - the analytical reasoning question. Zhong Wanjun tried various models and methods, "No matter what kind of ready-made model has no effect." ”
Exhausted but unable to solve the problem, Zhong Wanjun was immersed in the anxiety and uneasiness of finding a way and hitting a wall, and the idea of giving up this topic surged up in his heart countless times. But Mentor encouraged Zhong Wanjun to persevere, believing that they would be able to find a solution through the bottleneck period. In the end, Zhong Wanjun completely "started from scratch" to devise a method to "boil out the paper".
In order to solve the problem of logical reasoning in the entrance examination of the American University College of Justice, Zhong Wanjun and his collaborators proposed the LReasoner system, which generates answers to the question by recognizing logical symbols and expressions in the text. The LReasoner system consists of two specific parts: a logic-driven text enrichment framework and a logic-driven sample enhancement algorithm. This new approach sets a new record for ReClor, a dataset that focuses on the logical reasoning portion of the U.S. College of Justice Entrance Examination (LSAT), and exceeds human accuracy (click here to learn more about the research work).
Until recently, the LReasoner system has remained at number one in the official ranking of datasets.
The LReasoner system ranks first in the ReClor rankings Address: https://eval.ai/web/challenges/challenge-page/503/leaderboard/1347
Grow through trial and error and find your own scientific research model
One day in 2019, Zhong Wanjun, a first-year doctorate, opened his mailbox. The new news did not bring the long-awaited good luck, but sent her another rejection of the review opinion. This is the seventh time since her PhD that she has been rejected, and she has not yet voted for a single one.
This time working for AAAI is something she particularly enjoys. "Why is scientific research so difficult?" Zhong Wanjun, who has high requirements for himself, reads it repeatedly with the review opinion, collapses, sadness, self-denial... All kinds of tastes take turns to come to mind, and negative thoughts follow. Young ph.D. novices have repeatedly encountered setbacks at the beginning of their scientific research path, and it is inevitable that they will fall into confusion and self-doubt, "That was simply the most difficult period of my PhD." Zhong Wanjun said.
Mentor took everything into his eyes, told Zhong Wanjun about her past rejection experience, and encouraged her to adjust her mentality. Zhong Wanjun also gradually realized that he could not rush to achieve results, learned lessons from criticism, and gave himself more encouragement. When her mind became relaxed and peaceful, she was surprised to find that scientific research seemed to be less difficult and that progress was much faster than before. She also received her first top paper shortly after.
Scientific research always requires trial and error, and the path to the goal is not clear, which is one of the most difficult places in scientific research. Zhong Wanjun believes that in the process of this exploration, the role of the leader is particularly important. Zhong Wanjun's three mentors, Dr. Zhou Ming, former vice president of Microsoft Research Asia, Professor Yin Jian, vice dean of the School of Artificial Intelligence of Sun Yat-sen University, and Professor Wang Jiahai, director of the Institute of Big Data and Computational Intelligence of the School of Computer Science, have given her great help in scientific research direction guidance, paper submission and life path choice.
Whether it is choosing a research topic or formulating an experimental plan, personal blind exploration can easily fall into inefficiency and confusion, "Mentors and mentors can start from their rich scientific research experience to help me reach the goal efficiently and clearly, so as to avoid many detours." Zhong Wanjun recalls that from the initial confusion all the way to the present, I am extremely grateful for the role played by mentors and mentors, "just like a guiding light on the road of scientific research." ”
Over the past few years, Zhong Wanjun has gradually grown from a newcomer in the field who does not know much about deep learning algorithms, a scientific researcher who does not know how to write papers, and a novice in the laboratory who scratches his ears and scratches his cheeks in the face of design experiments, to a doctoral student who can be independent - relatively independently researching topics, determining tasks, building experimental benchmark models, completing experiments, and writing papers - Zhong Wanjun said that he learned too much from his mentors and mentors.
"I can feel that my research taste and my feelings about the entire field of research have changed qualitatively compared to before." Zhong Wanjun repeatedly stressed that the influence of good teachers on the taste of doctoral students is invisible and priceless. "Real frontiers, real challenges, really meaningful questions, really valuable approaches." Zhong Wanjun summed up the harvest of learning research taste from mentors and mentors, "Under their influence, I realized that having good research taste is even more important than the effort itself. ”
When failure changes from enemies to friends, setbacks change from suffering to wealth, Zhong Wanjun's mentality and scientific research rhythm are also changing. Once naïve and young whites, they often spent too long on a problem to drill the horns, or were too demanding and eager to achieve results, falling into a state of self-persecution. Now Zhong Wanjun has gradually explored a set of scientific research and life rhythms suitable for himself, "finding a peaceful, positive and stable state." "Every day I get up early to make breakfast, come to the workstation early to concentrate on work, go to exercise near noon, and continue to concentrate on work in the afternoon." From time-oriented to task-oriented, only require yourself to complete the planned tasks, no longer mandatory to work when you are not in a good state. ”
Zhong Wanjun unconsciously practiced Microsoft's "Work Hard, Play Harder" initiative, "Immerse yourself in it when you work, and when you feel tired, you will completely relax and rest seriously." When failure comes, you accept it calmly, draw strength from things that don't go well, and then "keep moving forward, without looking back or tangling."
Zhong Wanjun in life
Free and open, prosperous and growing
In the days of research and life at Microsoft Research Asia, Zhong Wanjun gained a free discussion of scientific research atmosphere and like-minded research partners.
"You can discuss issues with friends in all directions anytime, anywhere, and get along with Mentor very equally and easily." Zhong Wanjun said that since his PhD, many of his algorithm implementation methods were quietly born in the discussion, the visual computing group intern Meng Depu workstation is adjacent to her, Zhong Wanjun often discusses with Depp immediately after emerging new ideas, and the knowledge structure of the two sides complements each other, and has repeatedly promoted the realization of the experimental method of the paper.
The rich scientific research resources of Microsoft Research Asia also provide strong support and scientific guidance for students in the process of studying. From cutting-edge lectures, academic luncheons, to hackathons, to the "Teacher Shen Bring You Liver Papers" thesis reading workshop... Zhong Wanjun is active in various activities, "There are so many impressive activities! ”
She also has a very deep feeling for Microsoft's diverse and inclusive culture. By participating in Ada Workshop and Ada Dialogue, she met face-to-face with female researchers and engineers, and she realized that "the biggest problem for girls learning computers is not lack of ability, but lack of self-confidence." Girls are often instilled from a young age a negative view of 'not suitable for science and engineering', but I think that as long as they dare to break the prejudice and have a more confident mentality, girls can also perform very well in the computer field. ”
Wanjun Zhong attends the Ada Dialogue
In addition to academia, Zhong Wanjun is also actively expanding the possibilities of life. Every Wednesday board club event, weekend script killing, annual festival celebrations, and a variety of life knowledge and soft skills support courses... "Thank you very much microsoft research Asia for providing us with a natural and relaxed atmosphere, more fun than learning, and make many friends." "Participating in events with friends, trying out portraits, or researching cooking dishes at home on the weekends and inviting fellow students to try them out... Zhong Wanjun tried to find surprises in various directions, "The process of cooking is like doing experiments, all kinds of things are combined, but the result is unknown, which makes people feel novel." ”
Food made by Zhong Wanjun
Mentor's message
Nan Duan, Senior Research Fellow, Microsoft Research Asia
Over the past three years, Wanjun Zhong has done a series of excellent research work in the Natural Language Computing Group of Microsoft Research Asia and published many academic papers as the first author at top NLP/AI conferences, including ACLs and EMNLP. Especially in the field of machine reasoning, Wanjun has designed machine reasoning models for different scenarios based on different types of knowledge (text, logical rules, knowledge graphs, tables, etc.), and achieved the best results in the tasks including fact detection, logical reasoning and text-table question and answer.
I was deeply impressed by Wanjun's spirit of exploration of unknown problems and his quality of not being afraid of scientific research difficulties. I hope that Wanjun can continue to work hard in the field of NLP/AI, do more influential work, and become an outstanding female researcher in this field!