We aim to make AI that actually "thinks" like a human.
Here we introduce the AI research and development currently underway at NTT.
The field of AI has been attracting a lot of attention in recent years. We interviewed Kyosuke Nishida, a distinguished researcher engaged in AI research at NTT, about his approach to AI research and the future of AI research.
◆Profile: Paper Award at the 26th Annual Meeting of the Association for Natural Language Processing in 2020. Paper Award at the 25th Annual Meeting of the Association for Natural Language Processing in 2019. Best Paper Award at the 24th Annual Meeting of the Association for Natural Language Processing in 2018. DBSJ Kambayashi Young Researcher Award from the Database Society of Japan in 2017. IPSJ Yamashita SIG Research Award from the Information Processing Society of Japan in 2015. Part-time lecturer at Hosei University (since 2019).
I’m currently engaged in machine reading comprehension research. This technology understands text and answers questions correctly.
For example, let's say there is a sentence that explains the details and coverage of car insurance. It takes a lot of time to find information such as the hours when you can call for assistance and whether or not you can request a tow truck by reading from the explanatory text including detailed compensation details. So we use AI to do this. When a user asks a question, the AI explains the answer in an easy-to-understand manner based on the content of the text. For example, if the user asks “What hours can I call for a tow truck”, it answers “Tow trucks are available 24 hours a day, 365 days a year.”
AI has not yet reached a level where it can fully understand the meaning of text. But research has progressed significantly since the days when text was just treated as a set of words. Now, I am engaged in research to make AI correctly understand the meaning of words.
Today, AI can quickly read long texts on news sites and newspapers, which are very difficult to read, and can find answers to questions even in texts it has never seen during of its learning. We have made a lot of progress.
This technology is also expected to function as a chatbot or smart speaker, with AI responding like a real human. Until now, people used keywords to do searches, for example “Super Bowl location”. We aim to change this so that the user can ask “Where will the next Super Bowl be held?”, clearly conveying the intent of their search and asking a question with a natural sentence, to which the AI accurately provide pinpoint answers.
From the viewpoint of practical application, machine reading comprehension can be used for example, to accurately search for information from a large volume of internal documents. It is very difficult for an operator to fully understand a large volume of contracts and manuals. I believe that AI can assist with such difficult endeavors, or eventually take over the work entirely.
Many machine reading comprehension competitions have been held recently. At these competitions, AI is required to answer a question with natural expressions, rather than just picking out an appropriate sentence. The difficulty here lies in generating natural expressions. In my research, I use neural networks for natural language processing. When you enter information into a massive neural network, it neatly summarizes the sentence necessary required for a response and the content of the question to generate a reply sentence. When a simple answer is required, it can give just the answer part directly.
While the language comprehension ability of AI has improved considerably, there are still many texts in the real world that AI cannot understand. For example, if you go to a train station, you will see text on signboards and electronic bulletin boards all around you. Right now, it is difficult for AI to understand this text correctly. AI needs to understand the text layout and other such factors properly. It must understand all of the visible information, including drawings, text, and photographs.
Machine reading comprehension has great potential. My ambition is to make it correctly understand all visible text in the world, with a focus on language comprehension.
There have been major changes in the two and a half years since I started research on machine reading comprehension, especially in the past year. This is because of a major paradigm shift in natural language processing. Specifically, Google introduced a model called BERT, which dramatically increased the level of AI language comprehension compared to their previous model. The entire natural language processing field is undergoing major changes. This is a very interesting time, but competition is fierce.
Right now humans still have to adapt to AI, and various data needs to be shaped as inputs for AI. The scope of what AI can do is limited because it only solves specific tasks that have been learned in advance. I think that when AI can use the same information as humans to solve problems flexibly at the same level as humans, then humans and AI will be able to cooperate naturally.
The strength of our team. We were taking first place in the Microsoft competition from January to December of 2019. I think this came as the result of the strength of the natural language processing team at NTT, which was built up carefully over time.
Competing as a team has become especially important since research has entered the age of neural networks. The flow of research is becoming faster, so it is important to catch up with and surpass the world while also sharing information.
Some of the members of our team are from mathematics departments, or did not work on natural language processing when they were students. Up until a few years ago, I was working in the field of GPS log data mining and behavior modeling. I think a major advantage we have is that team members with various backgrounds are utilizing their individual strengths to do their best.
It is extremely difficult to do research alone. NTT also cooperates with universities and companies. From here on out, I think that we can make a bigger world, not only by focusing on our in-house team, but also by working with outsiders.
AI has existed for a long time as a research topic. However, since the speed of research is extremely fast these days, it is also a field in which you can catch up to the top level in short order if you try hard enough, if you enter after there is a change in the field. In that sense, you can still dream big, and I think there is a possibility that you can achieve great results as a researcher. This is a very interesting state of affairs.
Natural language processing is a field that has been attracting attention from all over the world, and I really enjoy doing this research.
I also think the environment is important to research. NTT has an excellent environment, and I hope we keep building solid achievements in the future.
Because business and research are changing drastically, I think it is very important to have an environment with many colleagues who inspire you.
There are various points to consider, such as having good researchers, motivated colleagues, and many opportunities for commercialization linked to business. But most importantly, you should enjoy doing research in an environment that suits you.