Discover Studying Abroad

9 Questions on Machine Learning and NLP with a Senior Engineer at Google

Natural Language Programming, along with Artificial Intelligence and Machine Learning, is a buzzword today. What are its challenges? What is it's future? We spoke to a senior engineer from Google.
BY Skendha Singh |   10-12-2021

Drishan Arora is a Senior Software Engineer who works for Google in California. He is a member of the Featured Snippets team and has helped build the question answering system underlying Google Search and Google Assistant.

BrainGain Magazine spoke to Drishan about his time at Columbia university, his work with Google, and his insight into both the problems surrounding NLP like fake news bots, and its potential in driving technical inclusion for the non-English speaking population of the world.

Below are edited excerpts from the conversation.

Machine Learning and NLP with a Senior Engineer at Google

  1. After graduating with a B.Tech from IIT Delhi in Electrical Engineering, what prompted your choice of Columbia University for your masters?

    Looking back, this was a pivotal decision in my life. At the time (about 8 years ago) Machine Learning and Data Science were emerging in a big way and although I had learned about them briefly, I enjoyed them a lot. Furthermore, Columbia University had invested significantly in this field and were building an exclusive data science division. They also had some of the premier professors in this field, who had joined their faculty recently, namely Michael Collins (with whom I work with now at Google) and David Blei. On a personal level, I also liked the idea of living in New York city and being in a diverse city.

  2. How was your experience of working as a Teaching Assistant while you were pursuing your masters?

    I was a TA for five courses in the two years I was at Columbia. Teaching is something I was always drawn to. Even though it started out as a good way to make some money to support my student life, being a TA taught me many vital and invaluable skills which are useful to me even today. Teaching AI techniques to over 150 CS graduate and PhD students helped me gain conviction in my own abilities and conceptual thinking. Like they say, if you want to learn something, then teach it.

    Eventually, I received a fellowship (named the CA fellowship) for two semesters which meant a full tuition waiver along with being paid for teaching. This financial assistance was a big relief in an otherwise expensive education program. Overall, the experience of working as a Teaching Assistant was indeed a very fulfilling one.

  3. Today you’re a Senior Software Engineer at Google. Please tell us a bit about your work with improving Google Search and the Google Assistant.

    At Google, I work in the Featured Snippets team. To provide more context, featured snippets are brief excerpts from a webpage that appear in Google's search results in order to quickly answer a user's question. For example, for user queries [such as why is the sky blue or who was struck by lightning the most] the resulting top passages that Google displays are Featured Snippets. This question answering system that I helped build is the underlying technology used in Google Search (and Google Assistant).

    I am very proud to have worked on this because open domain question answering is one of the most challenging problems in Computer Science and AI, which is also what made it super interesting to work on.

  4. Fake news and misinformation have led to major controversies for NLP – including the time when Microsoft launched the infamous Tay bot which had to be shut down after only 16 hours. How are researchers and engineers navigating those challenges?

    Fake news and misinformation are extremely important considerations especially at organizations like Google, where billions of people rely on you to find information on everything ranging from elections to vaccines. Consequently, we take this very seriously.

    There are a few fundamentals to keep in mind here. First, I believe that the most important part of building machine learning and AI models is not the algorithm you use, but rather how you evaluate them. In fact, the creation of evaluation metrics is often the hardest part to get right for most AI and NLP problems. Moreover, as important as it is to distinguish between a ‘good’ NLP algorithm and a ‘great’ NLP algorithm, it is even more essential to astutely look for edge cases to check where things can go wrong. For any user facing product, there is always rigorous testing before each launch.

    Second, some NLP technologies are inherently more “risky” than others. For example, the use cases where you ‘generate’ unconstrained information (such as Tay bot generating replies in a conversation) are most often riskier than email filters (deciding if an email belongs in one of three categories - primary, social, or promotions, based on its contents). Understanding these kinds of details is necessary to inform the process we then undertake to hedge for it.

  5. Whether or not we realize it – NLP is everywhere, from Grammarly to the chat bots. Can you share with us a few of the most significant applications of NLP today?

    This is indeed a very interesting question. It is hard for me to decide which is the most significant application but I can certainly talk about what I am personally most excited about.

    The way I think about it the field of NLP can be divided into two parts - one, where you analyze information (e.g., analyze if a customer review is positive or not; or whether a web result is a good fit for the search query) and the other is where you generate information (e.g., voice assistants or even translation).

    The latter is more thrilling for me. This is primarily because generation is fundamentally a harder NLP problem with a myriad of applications, which makes it more exciting to work on.

  6. We think of NLP mostly in the context of English. What about other languages? And Indian languages?

    Let me preface this by saying that it is my strong belief that India presents one of the most unique use cases for NLP because it has 122 major languages.

    Unfortunately, the majority of NLP research till date has primarily focused entirely on English and a few other languages. I am a huge proponent for Indian languages to find their much-needed place in NLP research as it is necessary for the technological inclusion of a vast population.

    I have personally led, and am currently leading, projects in this space. Aside from the social perspective, this makes sense from a technology perspective too. Working in different languages (and understanding NLP from a linguistic sense outside of English) can help improve NLP in English as well.

  7. Could you outline the role that NLP innovations can play in a country like India?

    Like I said earlier, India presents a very strong use case for NLP. For context, over 500 million people browse the web in non-English languages. Furthermore, that number is growing rapidly as more people come online. If we zoom out and consider the case of the internet, even though we think of it as open to everyone, most of the web is in English and a few other languages. This creates an ‘information’ divide in terms of what content the non-English language users (like Indian users) can access. This exists right down to necessary and universal features like input text tools such as keyboard support and spell checking, which do not exist for most Indian languages.

    At a higher level, the language you speak determines your access to information, education and the resulting economic opportunities available to you. NLP innovations can play a paramount role in increasing technological inclusion by magnifying the reach and impact of technology for a majority of Indians.

  8. Drishan Arora, Senior Software Engineer, Google in California
    Drishan Arora
    What is the future of NLP?

    I have been working in NLP and AI for over five years now and I have realized that the only constant in the world of technology is rapid change.

    Although the field of ML and NLP are ubiquitous and extensive, yet I would say they are still in the early stages. They are going to be exceedingly prominent in the next decade. However, what they look like in a few years would be very different from what they look like now.

    As an example, the NLP course that I took in Columbia (about 6 years ago) is almost completely outdated now. Working in this field has taught me that it is critical to keep on learning, innovating and evolving.

  9. You have a job that’s a dream for many across the globe. Any words of advice on making a career out of a passion?

    I am very humbled by this question. I realized at an early stage that there is a lot of uncertainty in life and that anything can happen. So, it is hard to decide "I want this exact job" and try to make it work backwards. Instead, I thought about what I wanted to do, and not who I wanted to be, and hoped for the best. This would be my advice to all the dreamers.



Can't Read  
Enter Above Code:


Sign Up for our newsletter

Sign Up for latest updates and Newsletter