Launch HN: Thematic (YC S17) Customer Feedback Analysis via NLP
We are one of the handful of companies that got into YC through the Startup School, and (I have to say) the only company that signed YC itself as a customer!
I have a PhD in NLP and ML and was consulting when two large media companies came to me with a problem: They collect tons of customer feedback in free text as part of their NPS surveys, but don’t have the time to sift through the responses.
This turned out to be common. Most companies collect feedback but, especially in large companies, nobody reads this data, and definitely not people who are in charge of strategy. Customers are screaming what’s wrong and what they want, but nobody is listening.
I tried a few open-source packages but found that none worked well. Developed on canonical text like news article or Wikipedia, they either failed to understand the variety of expressions, or were too hard to explain. I wrote a new approach capitalising on my PhD and new Deep Learning approaches. It's completely unsupervised: just needs raw data but, unlike topic modelling, produces clear and specific themes. My husband Nathan joined as a co-founder and for the next year we learned how to solve this problem in a way customer insights professionals find valuable.
Those media companies became customers and we quickly bootstrapped into a profitable startup. This is when Nathan signed up for YC’s Startup School. We grew 20% in those 10 weeks, loved the accountability and the focus. Our mentor suggested we apply for YC, which seemed like a crazy idea, but we gave it a go.
Fast-forward another 2 months, and we are just before Demo Day! Thematic grew 3x in that time, and we are working with brands like Vodafone, Air New Zealand, Stripe, Ableton, and Manpower Group.
Hope you found our story interesting, and happy to answer any questions.
32 comments
[ 3.9 ms ] story [ 81.9 ms ] threadWe show what areas to improve, not just how people feel (although that is important in deciding)
Would you mind please elaborating a little more on how you're thinking about theme-specific sentiment (for the NLP students here :) - is it along the lines of Socher's aspect-specific sentiment or something else?
You can check specific examples in our white paper: http://www.getthematic.com/net-promoter-score-verbatims/ (Don't need to download it, the bulk of the results is published on the page.)
In practice, some customers have read samples of responses and expect certain themes. Others have manually tagged them in the past. They compare their results with their fundings and can see straight away if our results are accurate and reliable.
I don't know much about NLP but are you only using unsupervised learning on the raw data? I would think you would need an NLP layer as well that sorts out basic synonymical issues, phrasing differences etc.?
However, we do have tools to review and adjust the results of Thematic by hand. For example, we have an internal drag and drop interface. Some customers really like to change the themes based on their view of the data. But it also helps to remove any inaccuracies, e.g. an incorrectly merged theme.
I'm also happy to make introductions if you're ever thinking about expanding up north to Canada.
Thank you so much for the offer of introductions. Definitely interested, as we sell internationally, not just in the US.
I for one really liked the demo and the blog - specifically, (a) I have great exemplars for what you mean by "theme", and (b) this post[1] shows great insights into your thinking about the problem faced by your customers
> Developed on canonical text like news article or Wikipedia, they either failed to understand the variety of expressions, or were too hard to explain.
It appears to me that the current methods and resulting tools are heavily dependent on the problem formulation (or domain in general). Moreover, no matter how fancy your technique is (or "how deep is your net"), the resulting model won't work unless you take specific steps to train it on data from the domain.
Yes, what I just said sounds borderline truism. However, I am more interested in discussing why it is so? Here's my initial thinking:
Let us look at (one of) the definition of Machine Learning, from Prof Tom Mitchell's textbook, "A computer program is said to learn from experience E with respect to some class of tasks T and performance measure P, if its performance at tasks in T, as measured by P, improves with experience E."
Here, experience E can be loosely considered as the amount of data you have for training - obviously, more data (i.e. training) should improve learning. However, the abstraction of T and P hides an important underlying problem of specification - or in other words, formulation of T (and E).
Thoughts?
> I wrote a new approach [capitalizing] on my PhD and new Deep Learning approaches.
I hope we get to see some of your insights in a paper or article (or blog post :)
[1] https://www.getthematic.com/post/visualizing-customer-feedba...
I like to think about ML in terms of how children learn language: through observation in their environment. The training data is a simulation of that environment.
Glad you liked the website and the blog post!
I work on the same, just for my own company to automate customer interaction (well, at least 99% of it).
I might have missed it on the website, how does pricing work?
Also, do you have any integrations with other tools like Intercom or Zendesk to ease data-sharing? A monthly insights report generated directly off of my main customer support tool can replace hours of manual work.
Do you guys see yourselves sticking to a model that spits out analysis, and let customers decide what insights to gain from the data? Or could there be a path where eventually it lets users take specific actions based on the data?
The more ways clients can hook the insights into their CRMs/workflow/user base, the more they can "operationalize it", as you eloquently put it, and make it a part of their workflows, the way the client you described does with their call center. I love it!
Example: Most popular topics/themes related to Vodafone: https://app.bloomberry.com/questions;q=vodafone
Surveys are the bread and butter for many market research companies. Most of the corporates / enterprises typically engage with smaller to larger (and many a times multiple) MR agencies. These MR companies can benefit from your service. To explore these companies, you can check out the MR members directory list from ESOMAR ,the voice of the MR (www.esomar.org), perhaps a membership / participating in their events may help you. Other site is agencyspotter.com
Explore publishing an article at the Greenbook blog run by Leonard Murphy which is very relevant to this case and he also runs the IIEX events globally ( http://iiex-na.insightinnovation.org/ ) where tools /services such as yours are very much the hot thing..
Check out Unilever Foundry https://foundry.unilever.com/ . You can sign up and explore if you can help solve some of their problems with your solution. They select and fund Pilot projects
Twitter hashtags to get your tool noted in the MR industry #mrx #newmr
Ads / promotions: check out http://newmr.org/ , http://www.greenbookblog.org/
Explore if you have complementary business synergies to present with https://www.zappistore.com/ (e.g your product could be a part of the Zappistore platform as an App.)
Best Wishes, N.Sankar https://www.linkedin.com/in/nsk007