By Kawshik Sarkar, Solutions Architect – AWS
By Shivank Bhardwaj, Technical Account Manager – AWS
By Piyush Patel, Chief Technology Officer – SuccessKPI

By Praphul Kumar, Chief Product Officer – SuccessKPI

SuccessKPI
SuccessKPI-APN-Blog-CTA-2024

In the current industry landscape, companies face significant hurdles due to lack of comprehensive sentiment analysis in call center data. This limitation poses a considerable barrier to effectively understanding and responding to customer emotions and concerns in real-time, which is vital for maintaining high standards of service quality.

The absence of in-depth sentiment analysis can lead to misinterpretation of customer feedback, resulting in missed opportunities to address grievances and improve satisfaction levels. Furthermore, this deficiency hampers the ability of businesses to accurately gauge overall customer satisfaction, a critical factor in formulating strategies that enhance customer experience.

In this post, we will explore how SuccessKPI leverages natural language understanding (NLU) and machine learning (ML) on a large dataset to predict sentiment. This approach enables customers to understand their sentiment towards consumed products or services.

SuccessKPI is an AWS Specialization Partner and AWS Marketplace Seller with the Machine Learning Competency. It supports sentiments of the following types: positive, neutral, negative, or mixed. Sentiment can be detected across multiple media types like voice, emails, customer support chat transcripts, social media comments, and reviews.

It helps businesses avoid personal bias associated with human reviewers by using artificial intelligence (AI)-based sentiment analysis tools. As a result, companies attain consistent and objective results when analyzing customers’ feedback and conversations.

How is Sentiment Determined?

The first step in sentiment detection is called tokenization. When a conversation is fed into the sentiment detection engine, it breaks the conversation into customer and sentiment channels. It then further breaks each of the channels into multiple turns, which is a group of phrases or a sentence.

In the second step, the sentiment engine prepares each turn by removing words that don’t add meaningful value to the phrase or sentence such as at, for, of, or other “stop” words.

The third step in this process is called stemming and lemmatization. Each word can have multiple versions, but they always have a single stem or base root. For example, the word “work” is the stem word for multiple words like “working,” “worked,” and” works.” So, it would look something like this:

working => work
worked => work
works => work

Lemmatization is very similar to stemming but has more context of the phrase and hence can differentiate between words which have different meaning based on part of the speech. For example, if we want lemmatize a sentence like “I am running and I usually use to run,” after lemmatization it would look like this:

I ---> I
am ---> be
running ---> run
and ---> and
I ---> I
usually ---> usually
use ---> use
to ---> to
run ---> run

The ML techniques and sentiment classification algorithms, such as neural networks and deep learning, are used to teach machines identify emotional sentiment from text. This process involves creating a sentiment analysis model and training it repeatedly on known data so it can guess the sentiment in unknown data with high accuracy. It uses phrases spoken, context of those phrases, the loudness, and raised voice as some of the factors driving sentiment.

How Are Sentiment Scores Interpreted?

Sentiment in conversations is recorded by assigning labels (negative, positive, neutral) and confidence scores. A typical sentiment output looks like this:

{
"SentimentScore": {
        "Mixed": 0.030585512690246105,
        "Positive": 0.94992071056365967,
        "Neutral": 0.0141543131828308,
        "Negative": 0.00893945890665054
    },
    "Sentiment": "POSITIVE",
    "LanguageCode": "en"
}

Machine learning in this case predicted there is less than 1% chance of negative sentiment and around 95% confidence of positive sentiment. The SuccessKPI algorithm picks the best prediction, which is positive in this case.

What Are Common Types of Sentiment Analysis?

  • Sentiment by channel: SuccessKPI currently supports sentiment by channel (customer or agent) and the confidence score of that sentiment determination.
  • Sentiment by time: SuccesKPI can detect sentiment at the speaker-sentence level. This allows customers to see the sentiment at the sentence level and understand how sentiment is changing throughout the conversation.
  • Sentiment by quarter: This allows customers to see the sentiment by quartile, so each conversation will be dynamically split into four quarters and then sentiment determined for each. This allows you to focus on a limited set of contacts to sample for quality assurance; for example, you can look at calls where you know the customer had a positive sentiment at the start but ended with a negative sentiment. That shows you they left the conversation unhappy about a topic.
  • Sentiment by entity: SuccessKPI publishes all of the entities detected in a conversation but there is no associated sentiment. This new capability allows customers to understand sentiment around a product or service. For example, in this feedback “the shoes are amazing but stores are dirty” there are two different sentiments expressed in the same sentence. SuccessKPI can distinguish the sentiment around entity “shoes” as positive and the sentiment around entity “store” as negative.

Examples of SuccessKPI Sentiment Dashboards

These dashboards offer essential insights across various dimensions, including sentiment analysis throughout diverse business units and stores. They also reveal sentiment trends across different service offerings, aiding in identifying and addressing potential risks.

Additionally, the dashboards provide geographical performance analysis to pinpoint regions facing risks, enabling strategic business pivots and informed decision-making.

Figure 1 – Dashboard representing sentiment across products and services.

What Are the Challenges with Sentiment Analysis?

Despite all of the advancements in natural language technology, machines continue to miss final details of human communication. Even if we leave the body language out of picture, verbal communications offer many nuances which can be hard to interpret.

Sarcasm. for example, is one of the situations in a conversation that’s difficult for machine to interpret. Imagine a customer that says, “Yeah, great. It took three weeks for my order to arrive.” Unless the computer analyzes the sentence with a complete understanding of the scenario, it will label the experience as positive based on the word “great.”

Negation is another example where machines could find it tricky to put sentiment correctly. For example, if a customer says “Oh yeah, it will take less than a week for you to go live with another vendor,” the machine could mark this as positive while the customer really meant that it would take more than a week.

Multipolarity is yet another aspect of speech that’s hard to interpret. Imagine a customer saying “The shoes were great, but the store was dirty.” In this sentence, two different sentiments were detected, and it gets hard for the machine to detect two different sentiments for different objects: shoes and store.

SuccessKPI offers ways to address this through entity-based sentiment where machines could detect sentiment around entities including products, names, and businesses.

How Do Businesses Benefit from Sentiment Analysis?

Understanding sentiment helps organizations understand the customer’s emotion or reaction towards a product or service. For example, say a global mobile carrier has 50 million subscribers and 30% of those are using Plan X. If 20% of Plan X subscribers have expressed negative sentiment across multiple media types, then subscribers unhappy with Plan X = 30% x 20% x 50M = 3M subscribers.

If the company can launch a campaign or fix the plan so it can save 2% of those subscribers, then subscribers saved = 2% of 3M = 60,000 subscribers. Now, say lifetime the value of each subscriber = $150. Then the total value generated for the business = $150 x 60,000 = $9M.

There are numerous business benefits to understanding sentiment analysis. Marketing teams can use sentiment to understand brand loyalty as well as the impact of a new marketing campaign. Often, product and marketing teams use this information to get alerted on a crisis before it gets out of control. Use cases are limitless from social media monitoring to understanding voice of the customer, customer care, and product feedback.

Conclusion

This post illustrated the critical role of SuccessKPI’s sentiment analysis in transforming customer experiences through advanced natural language understanding and machine learning technologies.

By accurately interpreting customer emotions and feedback across various channels, businesses can significantly enhance service quality and satisfaction. SuccessKPI, leveraging AWS’s powerful machine learning capabilities, offers a sophisticated, AI-driven approach to sentiment analysis, providing actionable insights that help businesses address customer needs proactively and improve overall satisfaction.

To learn more about integrating SuccessKPI into your customer service strategy, schedule a demo and see how sentiment analysis can elevate your customer experience.

.
SuccessKPI-APN-Blog-Connect-2024
.


SuccessKPI – AWS Partner Spotlight

SuccessKPI is an AWS Specialization Partner that supports sentiments of the following types: positive, neutral, negative, or mixed. Sentiment can be detected across multiple media types like voice, emails, customer support chat transcripts, social media comments, and reviews.

Contact SuccessKPI | Partner Overview | AWS Marketplace