FAQs: Sentiment analysis

The EU AI Act restricts the use of AI systems that infer emotions in workplace and educational settings, with exceptions only for medical or safety purposes (Chapter II, Art. 5(1)(f)). This restriction is driven by concerns about the scientific validity, reliability, and potential power imbalances of such AI systems in these contexts. Recital 44 of the Act emphasizes that AI systems using biometric data to infer emotions or intentions can be intrusive and lead to discriminatory outcomes, posing risks to individual rights and freedoms.

Genesys Cloud’s sentiment analysis operates differently and does not fall under this restriction. It analyzes customer interactions based on language used, assigning sentiment scores to phrases within interaction transcripts without using biometric data. For more information on this process, see Understand sentiment analysis.

Similarly, Genesys Cloud’s Agent Empathy Analysis evaluates the language used by customer service representatives to assess empathy and emotional intelligence during interactions. This analysis focuses on the dialogue within the transcripts and provides insights into the quality of customer care without relying on any biometric data. For more information about agent empathy analysis, see Understand agent empathy analysis.

For transcripts, sentiment analysis is always automatically performed after transcription is complete, the two are interconnected. It is not possible to turn off sentiment analysis for voice transcription without also turning off voice transcription.

For digital transcripts, sentiment analysis only occurs if the expected dialect is set to one specific supported language. For more information, see the Select one or more dialects for digital interactions section in the Speech and text analytics article.

For more information, see Understand sentiment analysis.

Sentiment analysis automatically highlights the exact moment when a customer conveys positive, neutral, or negative sentiment. 

The highlighted words and/or entire phrases become actionable business insights that can be used (among numerous use cases) to:

• Track customer perception.
• Identify and reward exceptional agent performance.
• Improve customer experience.
• Improve products and services.

Overall customer sentiment score

The customer sentiment score aims to capture the level of satisfaction experienced by the customer towards your company, products, or customer service. The customer sentiment score is computed by weighing each found sentiment event (positive or negative), with greater significance placed on the sentiment events that occurred towards the end of the interaction.  In essence, the customer sentiment score answers the question, did the customer leave the interaction happy, or dissatisfied?

How is the overall customer sentiment score calculated?

The overall customer sentiment score is represented with a value that ranges from -100 to +100, with -100 being extremely dissatisfied and +100 being extremely satisfied.

The overall customers sentiment score is computed by weighing each found sentiment event (positive or negative) based on their relative location in the interaction with a greater weight towards the end of the interaction.

The relative location is calculated by taking the index of the customer’s phrase (in which the sentiment event takes place), and dividing that index by the total number of phrases on the customer side.

In the following examples, assume that each call has 42 total phrases, 20 of which are spoken by the customer.

Let’s assume the following events were detected by customer sentiment analysis:

As a result, the overall sentiment score would be f((+1 x 0.50) + (+1 x 0.90)) = f(1.40), where f is the function that normalized the customer sentiment score in the range -100 to +100. The function f is defined as Normalized Overall Sentiment Score = 100 x tanh(0.75 x Overall Sentiment Score), giving a Normalized Overall Sentiment Score of 78.

Another example:

As a result, the overall sentiment score would be f((-1 x 0.10) + (+1 x 0.40) + (-1 x 0.85)) = f(-0.55), where f is the function that normalized the customer sentiment score in the range -100 to +100. The function f is defined as Normalized Overall Sentiment Score = 100 x tanh(0.75 x Overall Sentiment Score), giving a Normalized Overall Sentiment Score of -39.

For more information, see Understand sentiment analysis.

Currently, sentiment analysis is based solely on textual content that conveys specific customer sentiment (positive, negative, neutral).

For more information, see Understand sentiment analysis.

To run sentiment analysis for digital interactions (email, chats, and messages), you must set an expected dialect (language) in the speech and text analytics settings page.

Currently, English and Spanish dialects support sentiment analysis.

For more information, see Speech and text analytics settings.

The sentiment trend is determined by comparing the sentiment events found closer to the start of the interaction, to the sentiment events found closer to the end of the interaction. For this reason, the sentiment trend may be updated when additional follow ups occur within the same interaction. There is a minimum number of customer phrases required for the sentiment trend to be calculated, usually around 6 or more customer phrases are required.

Customer sentiment trend

Sentiment events are clustered into two groups:

• Sentiment events closer to the start of the interaction are defined as start-events.
• Sentiment events closer to the end of the interaction are defined as end-events.

sentimentTrend = (sentiment score of end-events – sentiment score of start-events) / 2

The sentiment trend is presented to the user as follows:

• Improving – If the trend score is +55 it is defined as improving.
• Slightly improving – If the trend score is between +20 and +55 it is defined as slightly improving.
• No change – If the trend score is between -20 – +20 it is defined as no change.
• Slightly declining – If the trend score is between -19 and -55 it is defined as slightly declining.
• Declining – If the trend score is less than -55 it is defined as declining.

For more information, see Understand sentiment analysis.