NLP applied to customer support, The most important skill for applied ML
I’m trying something new this week. This newsletter is called Applied NLP so I figured it’d be interesting to dive into specific business use cases for NLP. At the very least it’ll be fun for me because I pretty much spend most of my time thinking about how to improve business outcomes with NLP and speaking to non-engineering managers who are technology-agnostic and looking to improve their metrics.
There’s no shortage of company blogs or content that tout the technical brilliance of NLP. In my personal opinion, it comes across as having too many buzzwords and ambiguous in tying the technology to the bottom line. With the Use Case module, my hope is to cut through all of that.
Use Case - Self-Service Customer Support
Customer support managers closely monitor outcome metrics such as customer satisfaction scores. A key driver to achieving high scores is how fast the customer can get their questions or issues about the product answered.
According to Zendesk, 69% of customers want to resolve as many issues as possible on their own and 61% say resolving issues quickly is a top indicator of good customer experiences. However, the reality is that customers aren’t willing to put in that much effort to find the right FAQ article if they’re presented with multiple options in search results. Even if the answer is written down somewhere, customers will still file a customer support ticket and the agent will respond by just linking to the article which is a missed self-service opportunity.
This is a great use case for semantic search. Semantic search will not only return the correct article for the user to read but also the exact paragraph that contains their answer. It’s about removing the tiniest friction points in knowledge discovery so the user doesn’t feel the need to file a ticket. It’s a win-win situation because the customer gets their answer quickly and customer support agents aren’t burdened with repetitive requests.
If you’re interested in a guide for building semantic search systems are built, let me know in the replies!
Article of the Week
I was so excited to read this article from the Machine Learning team at Compass. It articulates what I’ve been thinking about which is that the key to building high-performing machine learning models comes down to understanding the “physics” of your business and the causal drivers of key outcomes.
Typically in machine learning courses, you learn that feature engineering or constructing the inputs to your model is the most important step in building a high-performing model. While that’s technically true, for pedagogical purposes you are often given a dataset to start with and derive features from. But in business settings, knowing what data to collect or generate in the first place is often the most important decision in the entire machine learning process. It’s also an expensive and time-consuming process where building the wrong dataset brings all downstream machine learning efforts to zero.
So how does one prevent this from happening? What can guide teams to collect the right data and engineer the right features?
Creating the right dataset involves formulating the right hypotheses of what causes the phenomenon you’re trying to predict. For example, the authors of the article walk through a visual conceptual model that outlines what variables impact someone’s likelihood to sell a house. The variables themselves have causal drivers so they end up with a conceptual model that’s layers deep.
Source: https://medium.com/compass-true-north/framing-feature-engineering-for-machine-learning-a-generative-model-of-home-likelihood-to-sell-f96f21d6d6d0?utm_campaign=Data_Elixir&utm_source=Data_Elixir_390
Going through this exercise and validating those assumptions guides the ML team to collect the right data which will then be transformed into high-quality features for a reliable model. In my opinion, this is the most important skill a Data Scientist or Machine Learning engineer can have. Similar to a Product Manager having product sense, the key skill is having the intuition of how data is generated within the context of the business and leveraging that to predict a useful outcome.