Deriving meaningful insight from 1000's of customer comments can be the difference between a loyal customer base, and a doomed company. But how is it possible to extract meaning from so much unstructured text? At Advize, we use a three-prong approach.
These three methods are categories, topics, and sentiment. Each has its own strengths, and together they provide highly accurate results capable of tracking known issues, identifying new trends, and uncovering how customer feel at about critical business concepts.
Categories are used to cluster conversations that cover specific business concepts. There are many different ways of speaking about the same concept, and any attempt accurately judge the importance of a business concept requires a model that will detect multiple ways of talking about a concept. For example, if the category was customer service, you could be speaking about agents, reps, wait times or issues resolution. To understand the impact of a concept like "customer service", it’s critical to bring all comments under one category for analysis.
We create these categories with a combination of deep learning and rules, that produce high recall and accuracy. Some categories can be applied to just about any business, but we also custom develop models for business to substantially increase the effectiveness of text analytics.
Topics are automatically generated based on the prominent concepts within a group of text and are aimed at taking an unbiased look at more granular things. Topics allow for a deeper understanding of what's going on inside a category, as well as identify new trends that are appearing outside of known categories. Topics and can be single words, bi-grams, or even short phrases.
Looking at the relationship between categories and topics (or two different topics) can draw meaningful insights about causality. An example in retail might be the category “returns" and topic "promo", or between two topics “promo” and "redemption”. Looking in at these relationship becomes even more interesting when adding in a third layer, sentiment.
While sentiment analysis is relatively straightforward to understand, it's application in text analytics varies greatly. We apply sentiment analysis at both the category and topic level to uncover which issues customers love and/or hate, and we apply for types of sentiment analysis “positive, negative, mixed and neutral.
For instance, if we go back to the NPS comment "The agent was friendly and helpful. However, your wait times are incredibly long”, we would tag three different things. The category “customer service” would receive a mixed sentiment rating, "agent" would earn a positive rating, and "wait times" a negative rating.
To learn more about our text analytics capabilities, you can schedule a demo here.