Optimizing Sentiment Classification Using NLP and Advanced Analytics
Companies across domains are constantly on the lookout for new ways to drive revenue. The insights provided from a customer’s experience with a brand or service have become critical for retaining and attracting new customers. Today, companies strive to ensure the consistent delivery of great customer experience through the meticulous analysis of customer feedback.
Organizations are using various tools to gain insights on providing better services. One of the most effective tools to gauge service efficiency is through sentiment analysis. Customer sentiment analysis determines the feelings of customers by deriving insights from an overview of the wider public opinion behind certain topics.
However, the sentiments expressed by the consumers online (in the form of feedbacks and reviews) are often incoherent, making it difficult to interpret the data. Fortunately, the right problem-solving mindset and advanced analytical toolset can be combined to overcome linguistic barriers and identify emotions in huge volumes of text in a matter of seconds.
One of the world’s leading Fortune 500 tech giant was looking to evaluate and improve their existing sentiment classifier tool. The client’s existing tool did not have a robust process for data cleaning and sentiment accuracy which resulted in the creation of inefficient models. Being unable to identify and prioritize the most critical problems in the current tool, the client approached Mu Sigma to help them revamp their sentiment classifier.
• Target: The subject on which the sentiment is being extracted
• Aspect: The feature of the subject.
• Fundamental Area: The customized cluster of features.
Eg: “ I like the acceleration of Mustang”
Sentiment: Positive Target: Mustang Aspect: acceleration Fundamental Area: Performance (derived from speed)
Following the methodologies of the Art of Problem Solving (AoPS) framework, we employed an outcome driven transformative approach (muOBI) to enhance the existing sentiment tool. Based on our insight into the current state, we began working towards the transformation of the tool.
Improvisation in New Sentiment Classifier
• RNN Model helped repunctuate the cleaned data to get more accurate text to sentence split
• NLP Stanza model helped capture the presence of multiple targets in a review statement
• A hybrid model involving Stanza NLP and deep learning classification models helped exhaustively identify the aspects in a review statement
• TASBA BERT model facilitated target level Sentiment Extraction
Upon implementing these steps, the new sentiment classifier tool was developed with the following advantages.
• Improved accuracy in Sentence Creation and Sentiment tagging
• Exhaustive and distinct tagging of Aspects/Keywords
• Fundamental Area Detection
• Secondary Target Tag – additional feature
The new sentiment classifier tool helped increased accuracy from ~80% to ~91%. In addition to this, Mu Sigma introduced two new features to the tool – Multi-Target Detection and Target Level Sentiment Extraction which lead to ~50% better text to sentence conversion and ~30% better aspect identification.
Over the past 15 years, Mu Sigma has solved some of the toughest business problems for the Fortune 500 using a unique combination of design thinking frameworks, plug-and-play innovation accelerators, and an army of agile decision scientists.
Want to know more?
Write to us at TheSherpa@mu-sigma.com.