This article is a brief summary of three presentations given at the monthly Data Meetup organized by BBVA Data & Analytics.
What happens when a comment is sent through a BBVA website, mobile application or email? Every day, BBVA receives thousands of emails from customers who want to ask questions or express their opinions about the quality of the services provided by the bank. BBVA customer service departments are increasingly using text analysis techniques to improve the response to this broad demand.
This article summarizes three text analytics methods developed by BBVA to process comments received through various channels and make that information actionable.
Predicting Comment Polarity
Predicting comment polarity is a technique designed to analyze the subjective component of a comment and more accurately predict the positive or negative outcome of an assessment. Based on Bo Pang and Lillian Lee’s proposal “A Sentimental Education: Sentiment Analysis Using Subjectivity Summarization Based on Minimum Cuts“, this technique allows BBVA customer service agents to sort and classify comments when they contact customers to offer new products.
The model was tested with 152,000 comments on the 250 top rated films in the IMDb film database. Given the premise that a film with a rating of less than six stars is considered to have a negative assessment, the model was capable of predicting whether the overall assessment was positive or negative by analyzing the text in the comments.
This method can be useful for classifying customers who have responded positively or negatively, and to avoid contacting anew those customers who responded negatively to a previous campaign and expressed a desire not to be contacted.
Automatic Comment Classification
Topic modeling is a text analytics technique capable of assigning tags to a set of documents. This technique complements the results of the most traditional classification and clustering techniques because it allows you to assign multiple tags to a document without supervision in order to identify common topics covered in a collection.
This technique, which was tested by analyzing over 370,000 comments about customer satisfaction, helped to automatically organize analyzed comments into 20 major themes such as “Number of Boxes”, “Bank Transfers”, “Customer Care” and “ATMs”.
Topic modeling can help to identify major underlying themes in customer feedback, because identified issues allow for the creation and refinement of a taxonomy of comments based on data that can later be used to speed up both manual annotation and automatic classification processes for comments.
BBVA’s Customer Solutions department has implemented a comprehensive methodology based on customer recommendations called the Net Recommendation Index (IReNe in Spanish), which allows them to determine satisfaction levels and make decisions to improve services and products.
On average, BBVA receives more than 80,000 responses per month in an unstructured format, and it is difficult for analysts to identify the reasons for customer dissatisfaction. Approximately one million responses were analyzed to uncover frequent topics of discontent, such as branch waiting times or the response time for questions.
By comparing the analysis results and IReNe’s indicators, analysts were able to determine action areas with high potential impact, and the Mexico office improved by 10 points in the customer satisfaction index.
Do you know of other examples of data analytics that may be useful for classifying customer feedback? Share them with our readers!