Park Byeongseon｜Interview

Please introduce yourself.

My name is Park Byeongseon and I joined the company as a new graduate in June 2019. I work at LINE KYOTO, which is a LINE's development center in Kyoto.

While searching for job opportunities, I heard about the opening of LINE KYOTO. I decided to apply for a long-term internship for engineering students, and that was what led me to join the company. During my internship, I was part of a development team responsible for speech recognition technology and handled tasks related to natural language processing. Then I turned my internship into a part-time job and was able to work on these tasks until I became a full-time employee. 

What surprised me the most was that, even as an intern and part-time worker, I was allowed to work on areas closely related to services. I only handled tasks that matched my work schedule, since I worked while attending university. But I was excited to be involved in important works and gained valuable work experience. To be honest, such experience at LINE deeply inspired me. Unlike other internships and part-time jobs that offered me limited duties, my role at LINE was completely different. When my manager at the time suggested the opportunity for me to continue full-time, I immediately decided to accept the offer. 

Could you please describe your main responsibilities and the specific workflow you follow?

After going to full-time work, I was assigned to a team developing speech synthesis at Data Labs, a specialized development organization for analyzing, researching, and applying data from all of the company's services. My main responsibility is language processing for speech synthesis. The technology is to convert text input into human voice. I work on areas to extract the necessary information for converting text into sound. Upon joining the company, I had multiple assignment options to choose from. Since there was a newly established team, I was assigned to my current position. I joined the team during the initial phase of defining its tasks and was able to participate in selecting our responsibilities with other team members, which helps me find fulfillment in my role. 

Let me briefly explain the language processing module that our team is developing with some examples.

The first process is text normalization. The process is for converting text with mixed formats including hiragana, katakana, kanji, numbers, English letters, etc. into a form that is easier to analyze. Japanese language has many elements, so it is arduous to analyze, but I really enjoy the challenges. An extreme example is, " yayyyyy★★★★★!!! It's half-past twelveee," this text can be transformed to "Yay! It is 12:30."

Next is pronunciation estimation. For hiragana and katakana, spelling matches pronunciation, but for other characters, we need to predict it. In Japanese, the same kanji or numbers can be pronounced completely differently depending on the context, so our goal is to correctly read these characters. For example, the text input "12時30分 (12:30)," will be processed by considering the combination of numbers and kanji, and the system outputs the correct pronunciation as "junijisanjuppun (12:30)."

Next step is accent estimation. This is a process that predicts natural intonation and accent in Japanese for a given text. For the input "12:30/junijisanjuppun," it will be processed as "12:30/juni*jisanju*ppun" (the asterisks indicate accent marks). This process is to phonetically convey the pitch-accent data for reproducing it as a sound.

 Although I explained the procedure very briefly, I hope it gives you a general idea of what it is like. Next, I would like to introduce the flow of my work.

My team focuses on R&D-oriented development, so we usually do not work in a short-cycle. First, we identify existing service problems and explore new approaches. After repeating a cycle of implementation, evaluation, and review, we then apply changes to the service. Let me elaborate using a case that I worked on in my first year on the job, where I developed a statistical accent estimation model for Japanese.

The first step is understanding the problems with existing service methods. Some outdated technologies and functions can be used for speech synthesis systems of services, but these are still in use. We study these areas in advance and come up with improvement ideas. In this case, the existing method used manually created rule-based accent estimation, which could not handle Japanese accent variants. We planned to enhance it with statistical approaches like machine learning.

Next step is exploring new methods, we survey similar techniques from research papers and case studies to develop concrete proposals for implementing our ideas. In this project, we investigated methods to apply sequence labeling systems, such as CRF and LSTM, to accent estimation and considered which method could achieve better results.

This initiates the phase of iterative implementation, assessment, and review. I'll introduce these steps one by one.

While implementing the plan refined by research and survey, we used CRF and LSTM for accent estimation. Specifically, we trained the model to predict accurate accents using converted input texts with a sequence of attributes like parts of speech, pronunciation, and pronunciation length. We used Python for development and first reproduced existing research, then improved accuracy by restructuring lexical attributes and modifying configuration of the training model.

When evaluating the implementation, we check its effectiveness and accuracy. In this step, we measured the prediction model's accuracy with our corpus and compared it to previous methods.

Then, by reviewing the model and evaluation results, we address any issues that may occur when applying it to the service. If improvements are needed, we repeat the implementation and evaluation cycle. In this case, we discussed within the team whether CRF or LSTM was more suitable for the service and identiﬁed bugs.

Now, next step is applying it to services. Once the specifications were finalized, we fine tuned the code and prepared to apply it to the service system, and then finally released it.

We work within a medium to long-term development cycle. I often work on multiple projects concurrently. When I developed the accent estimation model, I was also responsible for natural language processing tasks, including text normalization and pronunciation estimation. In addition, I also collaborate with my team members to improve morphological analysis and create automatically generating dictionaries. We stay up to date with the latest research trends, continuously explore and experiment new research themes and tasks.

Please share your experiences of what has been interesting or challenging in your field of work.

The exciting part is working with cutting-edge players who actively participate in top conferences. Whenever I learn about my colleagues' achievements, that motivates me to strive for higher levels. I think this is the most important factor of all. Moreover, we can work based on skill set and autonomy rather than experience. My job is never mundane, I find it satisfying and rewarding as I can manage development and scheduling based on my own decisions and interests. However, my challenge is that I still need to build my skills. I aim to elevate my skills to match my wonderful colleagues and contribute to the team while immersing myself in research and development.

As the fruit of my work, the paper I wrote about the Japanese accent estimation model was accepted by INTERSPEECH, the top international conference in speech processing. A demo page is also available, so if you are interested, please take a look.

I gained a more broad-minded approach compared to when I was a student. Since I have strongly realized that collaboration and discussion as a team yield better results than working alone on my own schedule. Because of this reason, I believe that the company respects each individual's personality which contributes to diverse perspectives. Therefore, any specific personality traits are not required. If you like open debates, I think the company is an enjoyable place to work for you.

Skill-wise, I think the knowledge of natural language processing and related expertise and development experience are essential. In my team, the skill sets of the whole-team is wide-ranged, so I concentrate on building expertise in the areas I am responsible for. I hope to deepen my machine learning knowledge and experience, as it would be fun to broaden my scope.

While in school, I recommend devoting yourself to research and writing papers in a related field. The work flow I explained is similar to the research procedures at universities and research institutes, so if you gain this experience at university or graduate school, I believe you can leverage it in your work.

What is the most rewarding part of joining the company?

As I mentioned earlier, I am delighted with the atmosphere and supportive working environment of both my team and the development organization. It may be obvious but working with inspiring people and being motivated to take on further challenges are very important aspects of work. Employees are friendly and communicate casually, and I really enjoy working with people who are striving for excellence.

I believe these factors also encourage me to grow further to achieve my career goals. Now I can clearly picture my next step by having role models in my team, who are successful both inside and outside the company.

Please share a message to our readers

I feel happy about working in such a stimulus-rich environment.

Being able to work with the most talented people in the industry and being part of such a team is a rare and exciting opportunity for any engineer. There is no job without stress, but I feel a strong sense of satisfaction that far exceeds stress, which I find wonderful. I think this is because I can fully realize my growth through learning. I would like to welcome you to experience it with me!