Physicist Willem Westra about how he applies physics in his daily work for Owlin
At Owlin, we couldn’t do what we are doing without the great work of all the different people working with us. In this interview series, The people of Owlin, we ask them about their daily work, background, and where they see Owlin going in the future. This month: Willem Westra, one of our Data Scientists, tells us what the product of Owlin has to do with the Theory of Relativity. Apparently, quite a lot!
You’re a physicist; how did you end up working for Owlin?
“One of Owlin’s early technical advisors, Pieter Scherpenhuijsen, convinced me to join the company. Because of his geophysics background, he knew exactly what to say to make me interested. Especially when he started talking about the mathematics behind Information Retrieval (IR), I got enthusiastic because it turns out that both in physics and IR so-called vectors play an essential role.”
Why is that?
“In mathematics and physics, a vector can describe for instance both the speed you are moving at and the direction in which you are heading. For example, in the case of three dimensions, you can move along a combination of the x, y, or z directions. Pieter showed me that search engines see documents as vectors in a very high dimensional space. Instead of x, y and z, every possible word corresponds to a new dimension.
When I learned this, I understood why Pieter saw a role for me at Owlin. As a physicist, I know a thing or two about vectors because one of my fields of interest was the general theory of relativity.”
What role do vectors play in the general theory of relativity?
“Albert Einstein developed the general theory of relativity, which explains the observed gravitational effect between masses as a consequence of their ability to warp spacetime. This explains why the earth is rotating around the sun: our planet simply travels as straight as possible through the curved space created by the sun. Both in the description of the path, as well as the shape of the curved space, vectors are heavily used.
While for general relativity theory, vectors are used to describe warped spaces, in the case of search engines, vectors are used to describe straight spaces. That got me excited, as it meant that warped space search engines had yet to be developed. I would say that even today this is an underexplored area of research”
And all of a sudden, you were the company’s first employee?
“Yes, this was a surprise to me. I just came back from working as a Postdoc in Iceland, and many people I knew were starting jobs at consultancy agencies. I contemplated joining an R&D department somewhere, but then Owlin came across and matched pretty well with my background.
Moreover, back then in 2012, it was not clear that working in data science was a good match for a physicist like me. Data science was not a thing. You can actually check it in Google Trends. Just search for ‘Data Scientist’.”
What did you like about working with Owlin?
“Working with Owlin is comparable to my Ph.D. and Postdoc because it allows me to work on a long-term project. And looking back, the thing I came up with back then is exactly what we are doing now.”
Has the product evolved a lot since you started?
“Yes, in many different ways, but the risk use case was pretty clear from the early days of Owlin. A big improvement is that while in the beginning, monitoring involved quite some manual labor for the end users, we were able to simplify this when we came up with the concept of ranking, which means Owlin simply sorts your portfolio of companies by their most relevant recent news activity. This is exactly what Owlin looks like today: the user will immediately see which companies most likely need to be checked.”
How would you describe Owlin?
“Owlin is Google News on steroids. We scrape the internet for news articles and transform that data into interfaces for (financial) organizations with large portfolios that want to monitor adverse media. We typically are a great addition to third party and vendor risk management, merchant monitoring, compliance and KYC business processes”
Ok, back to you being a physicist. How do you apply physics in your daily work?
“During the development of Owlin, machine learning became more and more of a (realistic) thing to consider. Machine learning is also all about vectors and their generalization called tensors. Let me explain what they are.
Tensors work like this. A vector is a ‘one dimensional tensor’ because it’s simply a row of numbers and the direction along the row is its single dimension. Matrices are ‘two dimensional tensors’ because matrices have two directions, one along the rows, and the other along the columns. If we go a step further, we have three axes: in that case, we have a cube with numbers on each side, producing a ‘three dimensional tensor’. This pattern continues to an arbitrary amount of axes giving us ‘n-dimensional tensors’ but we typically give up on visualizing these as that gets rather mind bending. Both deep learning and the general theory of relativity make heavy use of these n-dimensional tensors.”
Cool! But do you apply this knowledge in your daily work?
“Yes. For example, we currently work a lot with so-called `transformers’, a deep neural network that heavily uses complicated tensors. Also, I recently worked on an entity-linking project in which we researched the best way for Owlin to embrace this concept.”
Where is the data science team heading with Owlin?
“The focus is more and more on deep learning. One of the reasons for that is the great quality of tools and models that are becoming publicly available. A nice example is the sentence BERT library and its models, which we leverage to create similarity embeddings for use in document clustering.
Another reason why we are shifting more towards deep learning is that for some applications we have found ways to use it in a way that makes it explainable, a feature every data scientist should love.”
How do you use BERT for similarity embedding?
“An embedding is a vector in an n-dimensional space that can be used to describe, for example, the headline of a news article. The BERT model is basically a little machine that converts the words of the headline into such an embedding. To get the similarity between two headlines we simply calculate the angle between their corresponding vectors.”
Thank you, Willem!