Problem Description

A Fortune 500 client approached us about systematically keeping track of what their customers are expressing across various channels. For any organization it is very important to listen to the voice of the customer. However, for very large organizations this is challenging, because of the large volume of data their customers generate. Feedback from customers might appear on social media, through support chats, Q&A sections, reviews or emails. Our client was interested in particular in:

• Emerging topics
• Understanding when something is going well or not so well
• Tracking established topics of interest such as shipping
• Exploring the data as an aid in decision making. For example should we change our packaging? What are our customers saying about that?

A lot of progress has been made in Natural Language Processing (NLP). The main challenges of this use case were not the technicalities of NLP modeling, but rather how to extract relevant insights and how to make them actionable.

Reliancy Solution

Working with an internal analytics team we developed a comprehensive NLP pipeline and a dashboard applicaiton:

• NLP Pipeline To Pre-Process Text From Various Data Sources

  A proper pre-processing is crucial in NLP applications. This involves removing irrevant words and normalizing different variations of words that carry the same meaning.

• Ability To Detect Emerging Topics, Sentiment Analysis

  Trending topics are important when we do not know what to ask for. Specifically problems a client might not be aware of might arise as emerging topics. Sentiment analysis tells us how customers feel about a topic.

• No-Code Ability To Create A Detector For Topics of Interest

  Tracking topics of interest typically requires training data. We devised an approach that allows non-technical users to specify a list of keywords. Based on those keywords a topic detector is constructed and trained.

• Dashboard Application To Serve Insights and Explore Data

  We implemented a dashboard application to serve the most relevant insights to business stakeholders. The application allows for both emerging and tracked topics to be reviewed. A comparison is provided to what is considered "normal", so that peaks which are out of the ordinary are easily detected. A specialized utility to explore topic clusters was also implemented.

Impact

Working with an internal client we illustrated the utility of the approach:

• Understanding Where To Look

  Most companies already perform some kind of analysis of customer feedback. It may just not be systematic or scalable. For a topic of interest, providing a break-down of the sources where most comments are coming from turned out to be of value. When a topic is being assessed internal analysts know where to focus.

• Decision Support

  Working with an internal fulfilment team, we illustrated how our pipeline can be used to hone in on commets of interest when making business decisions.

• Detection of Problems

  While analyzing past data we found unexpected peaks. In some instances those were problems. For example multiple pleople complaining about images not rendering in search, which pointed to a technical problem customers were experiencing.