Visualizations to Make Sense of Your Customer Feedback
How leveraging NLP technology can help unveil what your customers are saying
Review sites and other sources of customer feedback offer companies overall ratings and computed scores to track their customers’ overall satisfaction. But these are essentially aggregated numbers that do not offer much help in identifying and quantifying specific strengths and weaknesses with regards to the customer experience.
On sites like Trustpilot, Tripadvisor, Amazon, consumers leave valuable data in the form of text letting businesses and other consumers know what they like and don’t like about a company, service or product. Whereas it would be impossible for humans to effectively summarize thousands of reviews and provide actionable insights, natural language processing and t ext mining techniques offer businesses quantitative tools to listen to their customers, ask the right questions and leverage text as data.
In this short article, we’ll explore four visualizations demonstrating how quantitative approaches to text can offer a clearer understanding of consumer opinions.
Beyond simple word frequencies
Word clouds are probably the most common way of visualising word frequencies in one or more texts. Usually the text is preprocessed such that only content words are present in the cloud. However, it may often be more useful to apply informed constraints when filtering the data. For instance, one can explore strings with two or more words, as well as target words belonging to specific parts of speech.
In this example, based on 5000 recent visitor reviews of The Little Mermaid statue in Copenhagen, I’ve made a word cloud consisting of two-word phrases from the review titles with adjectives followed by nouns. Applying sentiment analysis, phrases can be coloured according to the expressed attitude (red is negative, grey is neutral, blue is positive), though some ‘neutral’ words may well have been used to express negative opinions (e.g. ‘small’).
The statue IS quite small, and it is clear that some reviewers are rather unimpressed by The Little Mermaid. On the other hand, many reviewers seem to appreciate its cultural significance and it’s location by the water.
Co-occurring words
Another useful approach in quantifying and visualizing reviews is to measure how strongly content words are correlated, or how often they co-occur within reviews. In this case, 679 bad reviews of nike.com (n stars <= 2) are visualized in a word-correlation network.
In the periphery, we find Colin Kaepernick — the former NFL quarterback who kneeled during the national anthem, as well as a rather patriotic cluster of words. This is surely connected to the fact that Nike pulled a U.S.A-themed shoe after he campaigned against it. For the most part, bad reviews appear to highlight issues with long delivery times, refunds and customer service.
Uncovering topics
Topic modelling is a powerful tool in uncovering strengths and weaknesses with regards to the customer experience. Besides exploratory purposes, topic modelling has many other possible applications, such as input in supervised machine learning algorithms, customer segmentation, recommender systems, email routing and creating FAQs.
Topic modelling algorithms generally assume that texts are made up of a certain number of topics, and topics are made up of certain words. In the bar charts below, models of 1- and 5-star reviews (5.197 total) of booking.com with six topics are shown with the prevalence of each indicated by the bars. In addition the five most important words for each topic are displayed.
Topic models can sometimes be difficult to interpret and some topics may seem less coherent than others. Luckily, there are many preprocessing techniques and diagnostic indicators that may be applied to improve interpretability.
Likely and unlikely words
As a final example, we will explore which words tend to occur more in good vs. bad reviews. The data consist of adjectives and lemmatized nouns (i.e. words converted to dictionary forms) extracted from 2500 reviews of Amazon Echo Dot (3rd Gen). We compute log ratios of belonging to either category for each word and plot the words with the highest absolute ratios. This tells us which words are most likely to be found in either good or bad reviews.
Positive reviews seem to emphasize common everyday uses of the Echo Dot, whereas other reviewers appear to find the device more or less useless. It is also apparent that some buyers wanted to return the device. Further exploration of negative reviews would likely reveal why people wanted to return their Echo Dots.
Conclusion
There are countless ways in which NLP and text mining techniques can be combined and tailored towards answering important questions pertaining to the customer experience.
Many businesses have loads of valuable text data from surveys and online reviews just waiting to be explored beyond what is humanly possible. Whereas CX managers might have been more attentive to ratings while conducting small-scale qualitative analyses of the text material, the time is ripe for leveraging the rapidly maturing techniques for exploring massive quantities of precious feedback from customers.
Originally published at https://www.linkedin.com.
