Home> People and Culture> Interview Kei Uchiumi

Research and Development Group Researcher

Kei Uchiumi


Joined company in 2013
Portal Natural Language Processing Technology

“Solving problems that no one else has addressed” - that is what makes research fulfilling

Uchiumi has been researching natural language processing since before joining the company. At present, he is focused on the fundamental techniques of natural language processing with unsupervised learning. In order to achieve that research theme, Mr Uchiumi has experienced various positive “encounters”, adding to his technical background. We asked him about those encounters and past research results he has achieved.

The attraction of mathematics and the enjoyment of applying that knowledge daily

From the time I was in technical college, I’ve enjoyed solving complex mathematical problems. Perhaps because of that, writing code still brings me happiness to this day. For example, I try to make and implement algorithms in ways that are more efficient than the original paper I read.

The other thing is that the mathematical knowledge and thinking about algorithms that I have built to that point prove extremely useful. For example when I transferred from a technical school to a university, I did research into improving voice detection systems, together with a professor from the general laboratory of the National Institute of Advanced Industrial Science and Technology.

Encounters converge, leading to the world of natural language processing

When I first advanced to university, I didn’t yet have a clear idea of what I would study, so I started with being assigned that research by a professor. Meanwhile, in graduate school, I did research into language processing. Then, after obtaining my doctorate, I worked at a company that managed a major portal site, where I was assigned to the department of natural language processing. I had been wanting to research “information retrieval” and remember being a bit disappointed. But I ended up being highly influenced by the use of machine learning for natural language processing in the research that one of my senior colleagues was doing.


At the time, senior colleague was involved in using SVM (support vector machines: pattern recognition models built on supervised learning) and perceptron (a variety of neural network) to conduct natural language analysis. When I saw that, I felt the potential for using machine learning in natural language processing and began studying that on my own, applying my own programs. Later, as I learned about parsing and morphological analysis, I became completely absorbed in research about natural language processing.

Basic research into natural language processing is often compared to mixed martial arts. It requires more than knowledge of language alone, but also the knowledge necessary to apply mathematics, algorithms, and code. Without those, you cannot create programs that can be used. That was another attraction that made me strongly want to research the basics of language processing. As time went on, I was assigned to work with my senior colleague and we began releasing one research paper per year related to natural language processing.

Looking back, I was blessed with a number of beneficial encounters that have helped me grow, as a person and as a researcher.

Research into unsupervised learning of natural language processing is the first step toward holding a natural dialog with machines

Since joining DENSO IT Laboratory (hereafter, IT Lab), I’ve been allowed to carry out my wishes and continue my research into natural language processing. At present, the machine learning, such as neural networks, that is increasingly applied uses a high volume of data and requires a human presence to label correct answers. However, in the universe of natural language processing, data for learning is constantly created. For example, text data born from human behavior, such as blogs and twitter accounts, or comments on products, contains hidden training data. Speech recognition data concerning light conversations can also be used as training data. In this way, our conversations themselves can be used as learning data. Considering the massive volume of data available for natural language processing, there is no way humans can assign correct answers to each item.

In response, by researching unsupervised learning of natural language processing, I provide the program with a high amount of language data that is absent of correct answers, and then research how the program can actively look for hidden rules and structures. At present, I am testing how parsing analysis can be coded for application to unsupervised learning and we have gotten as far as morphological analysis and estimates.

Once this interview is over, from October 1, I will be shifting my place of work to my joint research partner, Nara Institute of Science and Technology, and intend to engage with parsing there. Later, I hope we can train machines to understand the meaning of sentences, but that will take a bit more time. Anyone who has used the smart speakers available for retail has surely noticed that they are not yet at the level of understanding the meaning of words. Rather, they search through key words and select the ones that appear most frequently. For example, if you ask “is there milk in the refrigerator”, the machine will respond “it’s at the supermarket”.

As basic research proceeds, the machines will likely learn to understand the meaning of recognized sentences and then become able to read between the lines. If that is accomplished, the machines will become able to fully understand human language and engage in conversation.

What interests me, however, is not so much using my research to create machines that can be conversation partners for humans. Instead, I would like to continue the basic research that will establish a foundation for unsupervised learning with natural language processing.

That’s because by deepening basic research, things will not stop with natural language processing, but will also extend to neural networks, where practical use is already proceeding, and tasks that until now could only be completed with a massive volume of labeled data will be completed using a lower amount, enabling solutions for tasks that are difficult for humans to address.

Working with difficult problems that are currently unsolved has the biggest impact

Since basic research takes a lot of time before it leads to business, most companies prefer to focus on research that meets short-term business goals. However, doing that leads merely to new arrangements of existing research, and doesn’t have a big impact on the world.

Luckily for me, the environment at IT Lab allows me to focus on basic research without concern for budget or time. That ability to continue with basic research is perhaps the biggest reason I chose to work at IT Lab. Thanks to that environment, I’m able to constantly ask myself “what are the problems that have yet to be solved” and concentrate my research in those areas.

Lately, a lot of researchers focus on problems that have already been solved and finding better ways to address them. I’m not so interested in that kind of research, and would like to continue answering problems that haven’t yet been solved and engage seriously with things that cannot yet be done. That’s much more interesting, isn’t it?

Anyone who is conscious of the same objectives can do research here while studying

Any students who wish to become researchers or who have a strong interest in research can achieve growth here, as I have, as long as they have the motivation, and I hope they will give IT Lab a try.

IT Lab is not a huge company. On top of that, there is barely anyone else who, like me, is studying natural language processing. Still, at IT Lab, I am encouraged to participate in academic conferences and can build broad connections through academic activities. In that sense, there are many opportunities to meet good teachers.

With regard to natural language processing, even if you are still lacking in knowledge and are interested in the research at IT Lab and in research into natural language processing, and are conscious of the same objectives, I can think we can work together to carry out long term research here even while you study. IT Lab even has a system for attending graduate school while continuing with research, so if you like, you can even pursue a doctorate.

To the list

Work at the lab Recruitment

Employment Information