WOMEN IN TECHNOLOGY SERIES
Charlene Chambliss: From Psychology to Natural Language Processing and Applied Research
An Interview with a Machine Learning Engineer at Primer.AI
Interest in data science has been exponentially increasing over the past decade, and more and more people are working towards making a career switch into the field. In 2020, articles and YouTube videos about transitioning into a career in data science abound. Yet, for a lot of people, many key questions about this switch still remain: How do you break into data science from a social science background? And what are some of the most important skills in fields like psychology that can be applied to data science?
Charlene Chambliss has an inspiring and non-traditional career path. Currently, she leverages state-of-the-art natural language processing to “build smarter tools for analyzing massive amounts of information.¹” In the past two years, she has written about NLP topics including BERT for named entity recognition, and word2vec for news headline analysis to name a few. However, before her current role as a machine learning engineer, she held roles in marketing, psychology, research, and interned as a data scientist in the skincare industry.
Amber: Could you tell us a bit about your background?
Charlene: Sure! I’ve had kind of an unusual path into data science, so I’ll start from the beginning and go into some detail to help illuminate what it took for me.
I grew up in a smallish agricultural town (Modesto, CA), where my dad worked at Safeway (still does!) and my mom was a stay-at-home mom. They really impressed upon me the importance of taking my education seriously, which was fine with me because I loved learning and I enjoyed making them proud of me.
Ever since I was little, I wanted to be a scientist. I loved tinkering and learning how things worked. My mom indulged my curiosity by taking me to the library (I would come home with a stack of like 12 books), having me help her in the kitchen (cooking = chemistry!), and getting me the occasional toy science kit.
That interest carried on through high school and into freshman year of college, where I had decided I wanted to study chemical engineering and become a flavor scientist, because chemistry was my favorite subject. I (adorably) thought that I would simply invent new flavors to make healthy food taste better, so people would have an easier time eating salads and vegetables, and thus be healthier overall. I hated eating salads and vegetables, so 17-year-old-me thought I was brilliant and that this was an amazing solution.
“Studying a social science is, in general, a great way to get used to dealing with muddy, hard-to-define questions, a skill that’s key to delivering data science work that decision-makers will actually feel comfortable using.”
I kept up my education focus and work ethic throughout high school, and made it into Stanford for undergrad. Frankly, that was pretty unexpected for me — I thought I would be going to UC Davis and maaaybe Berkeley if I was really lucky. Around half of folks who graduate from my high school don’t end up going to college at all, so even these felt like pretty high ambitions. Of my graduating class of 500 that year, I think only around 5 of us made it into “top schools” (Berkeley, Stanford, Harvard).
What I was really not expecting when I went to Stanford was the culture shock I was in for. The vast majority of students at Stanford come from upper-income backgrounds, with a median family income of $167,500. They are, by and large, the kinds of kids who have college-educated, professional parents, go to the best, most well-funded high school in town, and have paid tutors to help them out in any area they’re struggling with. Meanwhile, I grew up with a HH income around a quarter of that, and the level of preparation I received in some areas relative to my peers was reflective of that difference. (My parents and teachers were wonderful and had done their best, but there’s only so much one can do with limited resources.)
Suddenly, I found myself feeling very insecure about my abilities (particularly my aptitude for math and computer science) and was really questioning whether I measured up to the other students. I didn’t realize that our backgrounds had been so different, since no one goes around talking about that sort of thing, so I attributed differences in performance to my own lack of ability. I was also the only one from my high school who went to Stanford that year, so I didn’t know anyone when I got there and had no one to talk to about what I was experiencing. The feeling of being an impostor never really went away during my time at Stanford, but I did at least get better at faking-it-’til-I-made-it.
I did make it through Stanford, although I ended up not pursuing chemical engineering and also needed to take a year off after junior year to help with my parents’ divorce (my mom is disabled and needed help selling our family home and moving out). I graduated with a B.A. in Psychology in 2017 — first in my immediate family to get a 4-year degree — but I felt like I had made a lot of mistakes along the way due to a lack of guidance and role models. Even just searching for my first job proved difficult, because I could really only turn to the career center for advice on how to navigate the job market for “educated professionals.” The pamphlets and 30-minute consultations they could offer couldn’t really fill in all the gaps, but after a lot of research and attending career fairs, I was able to land a job doing social media marketing for a small agency.
Without going into too much detail about one’s financial and overall career prospects as a psychology major with only a Bachelor’s, it became clear to me over the course of my time in that job that I wasn’t going to get where I wanted to go career-wise unless I made a big change. So near the end of 2017, I decided I wanted to go into data science, specifically focusing on machine learning, and threw myself into GRE studies so I could get my applications in in time for Fall 2018 admissions. (I’ll go into more detail about why I chose data science, and NLP in particular, in the next section.)
I enrolled in my M.S. as planned, doing my coursework and studying as much as I could outside class, focusing especially on stats, linear algebra, Python, and machine learning. The degree coursework was all in R, so I learned Python entirely on my own using a combination of online classes and a massive 1500+ page textbook (Learning Python). Toward the end of my first year (spring 2019), I landed a data science internship at Curology and worked there through fall. Then, at the beginning of my second year, I partnered up with an amazing mentor, Nina Lopatina, through SharpestMinds, because I had decided I wanted to focus specifically on getting a role doing NLP. At the end of the 10-week mentorship, I started looking for jobs, and got an offer to join Primer full-time in December of 2019.
I would need to defer the last semester of my MS program to start full-time, which was a tough call, but the experience was more important to me, so I did. It turns out that that decision was frighteningly well-timed, because the COVID-19 pandemic decimated the recent grad job market only a few months later. I have classmates who are still struggling to find jobs, and I easily could have ended up in the same situation. I realize that I am very privileged that my roll of the dice worked out so well.
All in all, it took about 2 years to transition from marketing into a full-time machine learning engineer role, from a background of relatively little math and programming experience. (Prior to 2017, I had only taken single-variable calculus, basic/intro statistics, and one Java programming class.)
A: Before working in the data science industry, you studied psychology at Stanford. Could you tell us how your experience there influenced your career path into data science?
C: Studying a social science is, in general, a great way to get used to dealing with muddy, hard-to-define questions, a skill that’s key to delivering data science work that decision-makers will actually feel comfortable using. It’s sort of a Murphy’s Law mindset, as applied to experiment results: I’ve become extremely attentive to anything that could be “confounding” or otherwise influencing the results of my analysis, and I can call attention to potential caveats whenever appropriate. That way, the stakeholder can leverage their domain knowledge to decide whether they think those things do or don’t matter for our conclusions, and we can adjust the experiment/analysis accordingly.
In addition to that, I’ve probably spent a collective year and a half working in psychology labs to implement experiments. While there is a lot of “grunt work” involved in these sorts of positions, like data entry, you also get a front-row seat to how scientific studies actually happen, from data collection to the final statistical analyses, and you get to participate in some of the decisions that are made along the way. This prepared me quite well for data science workflows, as well as giving me some practical skills (like working with spreadsheets) and a can-do attitude that would be helpful later.
A: Previously, you interned at Curology as a data science intern. Could you discuss how data science looks like in the skincare industry? What types of questions did you and your team seek to answer? And what are some of the most interesting projects you worked on in your time at Curology?
C: I think my experience at Curology was a good example of how data science looks in a D2C (direct-to-consumer) business in general, especially in a startup context. It is often the case that the first thing consumer-focused businesses need data-wise (after data engineers, of course) is really just a lot of descriptive statistics, often known as “consumer insights.”
Since I was embedded in the user acquisition department, I was especially focused on answering questions that would help us make better marketing decisions across the many different acquisition channels. 80% of the time, I was writing SQL against our data warehouse to better understand the behavior of different customer segments and track how that behavior trended over time, and turning those findings into interpretable dashboards for use by the rest of the team. The other 20% of the time, I used Python to analyze and visualize customers’ survey responses to better understand what they liked and needed from Curology.
So a few of the questions I got to ask and answer were:
- How do customers’ skincare goals vary based on their demographics (gender, age, etc.)? What is most important to each segment of customers, and how can we make sure we serve each of their needs well?
- Which of our channels have had the “stickiest” customers, i.e. customers who have tended to stay with us the longest? Do any other behaviors or preferences correlate with subscription length?
- Can we build a model that will leverage the historical data we have on customer behavior to predict customer lifetime value (LTV) at time of signup? (This is actually very hard when your customer base is growing quickly, due to sampling considerations!)
I learned a ton. Doing data analysis with SQL doesn’t just help you learn SQL; it actually helps you think analytically, as cliche as that sounds. You first have to learn to translate someone’s natural-language question about customers into the appropriate metrics (where those metrics will often have different filter conditions and assumptions, depending on the intended use-case!), then ALSO learn how to actually execute that in a mathematically and technically correct way using SQL code. Sometimes you will even have to make sure that you are using the correct tables/data, because tables get deprecated, not all data makes it into the table due to bugs in the pipeline, or X metric only started being tracked 6 months ago, etc. There are many practical considerations you need to keep in the back of your mind when doing this kind of work. Doing rock-solid data analysis is just as challenging as machine learning IMO, albeit sometimes for different reasons.
A: Did you always know that working in data science was what you wanted to do? What inspired you to pursue a career in natural language processing? And, could you tell us a bit about your work on the Primer.Ai applied research team looks like?
C: Not at all! I don’t think many of us who work in data science today could have anticipated the rise of this field. I didn’t even really know about the widespread use of applied statistics in the private sector until my senior year of undergrad.
When I graduated with my B.A. and started my first job in marketing, I figured out that that wasn’t the right fit for me relatively quickly. I started researching my alternatives to see if there might be a career I could transition into that would be better-suited to my personality and values (and frankly, better-paying, as the entry-level marketing salary was only enough to live paycheck-to-paycheck in the Bay Area).
After a few months of digging, I landed on data science. It was intellectually challenging work, poised to make an enormous impact both economically and in society at large. Not only that, but I noticed that people in data science careers often cared more deeply about ethics than I had seen elsewhere. To see that people in the field genuinely cared about how their work would impact people really spoke to me, and is what ultimately helped me decide on making the transition.
That said, I was still unsure, because I had had negative experiences with math and computer science in undergrad, and I wasn’t sure that I could hack it (haha). In my first quarter at Stanford, I got the worst grades I had ever received in my life in a calculus course and a CS course, which caused me to seriously question whether I was cut out for those kinds of subjects. When I started this journey, I had to convince myself that I could succeed by using objective measurements instead of my own feelings: “well I scored X on the SAT, and the average score for CS majors on the SAT was Y (where X > Y), so I should be able to learn the math and other material just as well as other folks in this field…”
“I love NLP because I can contribute directly to helping people cut through the noise and get down to what they need to know in order to live their lives and do their work more effectively.”
I later made the connection during my M.S. in stats that the primary reasons for my underperformance as an undergrad were a lack of good study habits and a lack of interest in math as a subject. In high school I could get away with waiting until the night before to study for the test, and never reading the textbook outside of class, but that was no longer the case at Stanford. I improved my study habits over time, and by the time I took 2 statistics courses in senior year, I was able to ace them both. Similarly, once I started learning about some of the fascinating and unexpected ways that math and stats are being applied to the real world via data science, my mind really awoke to the benefits of math, and suddenly the motivation to master it was there. I got straight A’s in my M.S. coursework for all 3 semesters that I was enrolled.
It is still a little crazy to think that just a few years ago I truly disliked both math and programming, yet here I am now, using them both every day and genuinely enjoying it. I really want to emphasize how important it is not to put yourself into a “math person”/”not a math person” box, and the same goes for programming. Both skillsets are simply tools, and these tools have incredible power to make you more effective at any other area or interest you care about making a difference in, whether that’s art, law, a social science, or a more traditional synergy like engineering. If you can push through those early feelings of resistance and intimidation, there are wonderful feelings of competence and accomplishment waiting for you once you’re able to start using these tools for the things you care about.
As for why I chose natural language processing (NLP) in particular, there are a few reasons. On a career level, I saw the NLP community as more welcoming to people from unconventional backgrounds, relative to an area like computer vision where I was really only seeing people from CS, math, physics, and electrical engineering backgrounds. On a more personal and interests-based level, I see NLP as the field best suited to helping solve the problem of information overload. There is an endless amount of information to consume, which is contributing to a heightened level of stress for everyone, as well as impairing the productivity of people in knowledge work careers. I love NLP because I can contribute directly to helping people cut through the noise and get down to what they need to know in order to live their lives, make informed decisions, and do their work more effectively.
My work at Primer is directly relevant to the problem of information overload. At Primer, we’re leveraging powerful, cutting-edge NLP models to extract structured information from noisy, unstructured text data. This helps our customers get at the information they need much faster than having individual humans poring over the data themselves. Some analysts are working 12 hour days simply because they have no way of quickly reading and digesting the deluge of information they’re responsible for staying up-to-date with, and we want to change that.
My team, Applied Research, is tasked with training, testing, and making deep learning models available for Primer’s products, then integrating those models into our data pipeline or exposing them for use via an API. We also create reusable scripts and resources that allow people to train their own models on their own data. The work involves not just model experiments and engineering, but also plenty of collaboration with other teams that work more directly on our products and infrastructure.
In terms of the week-to-week, I’d say half the time goes to writing code for model training/evaluation, data preprocessing, and other typical machine learning tasks, and the other half goes toward communicating about the work: discussing plans, specifications, and progress with product managers, working with our data labeling team to create datasets for new and existing tasks, as well as presenting to the company at large about new developments and improvements of our models.
A: During your fellowship at SharpestMinds, you developed a toolkit for training “BERT-based named entity recognition models” for an error analysis frontend in Russian to English machine translation. Could you describe your project more detail and share what your three most important takeaways are?
C: The TL;DR of the project is that my mentor needed a way to train BERT models to do named-entity recognition in Russian and English, and my task was to learn how to do NER with BERT in PyTorch, then build out the entire pipeline in the form of a git repo that could be cloned and run locally.
The resulting trained models could then be used to highlight entities, such as people, places, and organizations, in a user interface, where translators would identify whether a separate model (a Russian-to-English translation model) had made mistakes when translating names from Russian to English. I thought this — using the models to create more powerful, human-friendly software — was super cool, and this project really served to develop my intrigue for building ML-powered tools and interfaces.
At the time that I did this, there was exactly one blog post about using BERT to do NER, and the code didn’t work out-of-the-box for me, so needless to say: there was a lot to figure out along the way! (Regardless, props to the author, Tobias Sterbak, for a very useful post; without it, it would have taken me a lot longer to get started.)
For anyone who would like more details, I wrote a 2-part series for In-Q-Tel Labs’ blog about the project as I was wrapping it up. Here are Part 1 and Part 2 of the series, and the repo can be found here.
C: What advice would you have for other women who are looking to enter the field?
C: If you’re still in school (especially undergrad), you have 3 good options: study computer science and get a broad CS education, study a quantitative subject like applied statistics, economics, or engineering and combine with CS coursework, or study a qualitative subject while teaching yourself how to apply DS/ML to your field. Get research experience, especially if you want to pursue a MS or Ph.D., and if you think you prefer industry, get industry experience (have an internship every summer, maybe even work part-time during the school year). Whichever path you take, be aware that some conceptual understanding of the math behind DS/ML (statistical estimation and probability, linear algebra, and calculus) and good programming skills are still required to be successful in most roles.
If you are a career-changer already out of school, study the career paths of people who are already in the industry. Try to pay the most attention to people whose backgrounds are similar to yours: for example, if you’re coming from a “non-technical” field that doesn’t involve much math or programming, take note of how other people transitioned in from non-technical fields. Figure out what they needed to do in order to prove that they had sufficient technical skills. Reach out to those people, see if you can get 30 minutes of their time for a phone call, and ask them specific, focused questions on what you would need to do to become hirable for the kinds of roles you’re interested in.
Career changers should also strongly consider going through a mentorship program such as SharpestMinds, if you’re transitioning from an unrelated background and need help designing and scoping an impressive, professional-quality data science project and preparing for interviews. If you are coming from a lower-paid field as I was, the income-share agreement is a lifesaver since you don’t have to pay anything until you actually get hired in a data science role.
Also, Vicki Boykis’ article Data Science Is Different Now (written last year) is required reading for any aspiring data scientist. I don’t agree that everyone who takes a Coursera course or even a bootcamp is necessarily qualified for entry-level data science, but it absolutely is the case that competition for these roles is fierce, and you will need to do something to differentiate yourself from the many aspirants. As Vicki suggests, taking an adjacent role in general software engineering or data analysis first can be extremely useful for building skills and getting your foot in the door.
A: How can our readers connect with you and get involved with your projects?
C: You can follow me on Twitter, connect with me on LinkedIn, or just send me an email. I’m pretty heads-down on my work at Primer right now, but if I start any side projects in the future, I’ll be sure to share!
In addition to innovating in the NLP space as a machine learning engineer, Charlene has also been an active member in the SharpestMinds community. Her career path and multi-disciplinary experience focused on social science and technology continues to empower women technologists to innovate in data science. Today, Charlene is solving the toughest problems in NLG and NLU, thus giving people the power to cut through the noise and understand the world at scale. As well as being a role model for data scientists who have non-traditional career paths, Charlene inspires women from diverse backgrounds to democratize the field of NLP.
Special thanks to Charlene Chambliss for allowing me to interview her for this series, and a huge shout out to the TDS Editorial Team for supporting this project.
Do you know an inspiring woman in tech who you would like featured in this series? Are you working on any cool data science and tech projects that you’d like me to write about? Feel free to email me at angelamarieteng@gmail.com for comments and suggestions. Thanks for reading!
[1] Additional information obtained from LinkedIn, available upon request from the author.
References:
“Economic Diversity and Student Outcomes at Stanford — The New York Times.” n.d. Accessed August 15, 2020. https://www.nytimes.com/interactive/projects/college-mobility/stanford-university.
“How to Fine-Tune BERT for Named Entity Recognition | by Charlene Chambliss | Gab41.” n.d. Accessed August 15, 2020. https://gab41.lab41.org/how-to-fine-tune-bert-for-named-entity-recognition-2257b5e5ce7e.
“Charlene Chambliss (@blissfulchar) Articles — Medium.” n.d. Accessed August 15, 2020. https://medium.com/@blissfulchar.
“Using Word2vec to Analyze News Headlines and Predict Article Success | by Charlene Chambliss | Towards Data Science.” n.d. Accessed August 15, 2020. https://towardsdatascience.com/using-word2vec-to-analyze-news-headlines-and-predict-article-success-cdeda5f14751.
“(19) Charlene Chambliss | LinkedIn.” n.d. Accessed August 15, 2020. https://www.linkedin.com/in/charlenechambliss/.
“Primer | Press.” n.d. Accessed August 15, 2020. https://primer.ai/press/.
