Customer Sentiment Analysis for contact centres

What is sentiment analysis?

One specific use of natural language processing (NLP) that has found a home in customer service is “sentiment analysis”. Sentiment analysis (sometimes called opinion mining or emotion AI) is the identification, extraction, and quantification of the positivity or negativity of a piece of text with statistical techniques such as machine learning. 

In customer service settings, sentiment analysis is very widely applied to feedback received from customers in customer contact centres, often from post-contact surveys or submitted product reviews.

Sentiment analysis is particularly attractive because it changes unstructured, messy data in customer conversations and reviews into structured, reportable, and actionable data that can be used to inform decision making internally, quantify issue severity, and to improve the customer experience. 

Analysis of text content in this way helps companies to understand what is driving frustration or satisfaction in their customers. When the drivers of particular customer emotions are known, specific steps can be taken to improve the customer experience, which translates directly to improved customer loyalty, and ultimately, increases in revenue.

Sentiment analysis is particularly attractive because it changes unstructured, messy data in customer conversations and reviews into structured, reportable, and actionable data

Sentiment analysis techniques

The most basic sentiment analysis techniques, are “word list” techniques, which use lists of positive and negative words (e.g. bad = negative, happy = positive) and score content simply by counting the number of occurrences of each. Word list approaches provide a baseline score, but they are brittle. For example, the phrase “I’m really not happy or satisfied about your new sale prices” technically has more positive words than negative.

Newer, more advanced approaches will parse sentences using NLP techniques to unravel sentence structure and machine-learning models to cater for slang or sarcasm. The best systems can detect tone in voice conversations, take word order into account, and can decompose sentences into different entities driving sentiment polarity.  The most accurate results for sentiment scoring are achieved by training custom models for your own customer conversations and vocabulary using a manually created training data set, but this process can be typically labour intensive and expensive.

Off-the-shelf sentiment models

There are a wide range of sentiment analysis APIs and machine learning models that can be downloaded or accessed online, examples including large cloud providers such as Amazon Comprehend, Google Cloud Natural Language, and Azure Text Analytics, as well as do-it-yourself approaches with Python Flair, NLTK, and HuggingFace Models.

Unfortunately, many off-the-shelf sentiment analysis models that are available online or through programming frameworks are not fine-tuned for the customer service use case. 

Many models are trained on movie and product reviews that can be found online with associated scores, enabling an easier training cycle. At EdgeTier, we’ve found the best results for sentiment and emotion detection for customer service can only be obtained by training custom models through manually labelled data. While laborious, the accuracy of the resulting model is vastly superior.

Beyond a sentiment score alone (typically “positive” or “negative”), an extension of sentiment analysis is “emotion detection”, which breaks down text content into the underlying emotion of the speaker(s). At EdgeTier, we have developed customer-service-specific models for “frustration”, “delight”, and “gratitude”, which are useful emotions for agent review and product feedback. Models can technically be built for any emotion that can be labelled or determined from text content. 

Uses of sentiment analysis in Customer Care

If you have access to an accurate sentiment analysis tool, it can be applied on different sources of data from your customer service center:

• Conversation Transcripts: Sentiment analysis on customer conversations such as chat transcripts, emails, and phone transcripts (transcribed using speech-to-text technology) gives an insight into how satisfied or dissatisfied customers are when speaking with your agents.
• Agent Behaviour: Turning the technology to focus on agent utterances across all channels helps to identify issues with agent behaviour and variances in tone-of-voice.
• Social Media Monitoring: Whilst not directly related to customer service, sentiment analysis is often applied to external data sources such as social media to aid marketing departments to gather brand opinions. For customer service, insight on any real-time issues driving negative sentiment may be useful to inform agents when customers do contact.
• Survey Submission Verbatim: Customer survey results are a prime source of sentiment information for companies, where users are specifically asked about their opinion on issues or product features. Sentiment analysis results allow analysts to isolate the issues that are driving the most negative sentiment and quantify the scale of those issues.
• Anomaly Severity Assessment: Sentiment can be used to quantify the scale or severity of anomalies or issues detected in the customer service team. If an issue occurs but customers aren’t disgruntled, perhaps something else can take priority.

Sentiment Analysis Project outcomes

When you have a sentiment analysis system up and running, what should you expect the impact to be for the customer contact centre?

Ideal Outcomes

• Improved customer satisfaction and retention though actions taken on sentiment results. E.g. clearly identifying the drivers of poor customer sentiment and taking actions to repair or fix the customer experience.
• The ability to dissect customer verbatim and use the insights to inform and drive operational decisions across the company.
• Adoption of sentiment results across the company as key part of the “voice of the customer”.
• Time savings in manual reviews of customer text fields from surveys and customer chat and email transcripts.
• Fast identification of customer service agents that are communicating outside of brand guidelines with incorrect tone or attitude.

Potential Pitfalls

• Siloed adoption by only small teams within customer service
• Inaccurate detection results through the use of non-state-of-the-art models
• Using an off-the-shelf model on text content that is too dissimilar to the training content