Viewing sentiment in reports

Aggregated sentiment information highlights the aspects of the customer experience that have a positive or negative impact on how customers feel. For example, in the following report, 69% of phrases about channel choice are negative. These phrases are also associated with lower NPS scores, an indication that widespread dissatisfaction with channel choice might be pushing overall NPS downward.

Phrase-level sentiment information reveals the sentiment class associated with an individual phrase — a short sentence, a clause, or a group of words that convey a meaning. In Medallia Text Analytics, you can view the sentiment classes of the phrases that have all been categorized into the same topic. In this example, the highlighted phrases are associated with the same topic. The red phrase is negative, and the green phrase is positive:

Users can see the result of sentiment analysis in reporting modules, such as the TA Topics/Themes Explorer module. The following image shows a phrase list in the TA Topics/Themes Explorer module with a positive comment expanded. The highlighted text shows the text Medallia AI used to determine that the feedback was positive.

You can combine sentiment analysis and suggested actions highlighting in a Responses Feed module. With the module opened in Admin Suite, select whether you want to display topics or themes, and then select how you want to filter by those topics or themes. For example, the following image shows properties configured to show topics, and to allow the user to select topics from the control panel:

Ensure that the Enable TA Highlighting and Enable Suggested Actions Highlighting properties are turned on, as shown in the following image:

Sentiment scores

In Text Analytics reporting modules you can select from a variety of scores to display, including the following scores that factor sentiment into the calculation. Sentiment scores are especially useful when customer feedback is either fully or partially from sources that do not provide a traditional key metric (such as NPS). For example, sentiment scores make it easier to understand feedback received through social media.

• Net Sentiment — The percentage of records that have Positive or Strongly positive sentiment minus the percentage of records that have Negative or Strongly negative sentiment. Use this score to better understand how customers feel about a particular topic or theme.
• Impact on Net Sentiment — The impact of a specific topic or theme on the net sentiment score. Use this score to better understand which topics or themes should be prioritized as areas for improvement.

For more information about sentiment calculation, see Reporting metrics and calculations. For more information about configuring Text Analytics report modules, see Reporting.

Sentiment classes

Sentiment model development and availability

Sentiment models are created and maintained by the Medallia research team. The process of creating a new sentiment model begins with a data set that includes at least 400,000 sentences used to train the sentiment model. Each sentence is labeled with a sentiment class by two or more people to help ensure that inaccurate sentiment annotations are not introduced.

The Medallia research team then applies state-of-the-art machine learning techniques to the training data (the annotated sentences) to discover features and patterns inside the text, and to find boundaries that separate different sentiment classes.

Sentiment-model learning techniques are similar to techniques humans often use: observation and experience. For example, after a child sees an airplane fly for the first time the child can draw on that experience to recognize any airplane as an object capable of flight. Sentiment models learn to classify sentiment in a similar way. The sentiment model’s experiences are the sentences manually labeled with sentiment, while the sentiment model’s observations are the features and patterns in those sentences.

The Medallia research team selects a set of features that optimize overall sentiment accuracy, and then finalizes the initial sentiment model. After building the initial model, the team conducts user acceptance tests to evaluate how accurately the model meets accuracy requirements, and iterates the model until accuracy is at an acceptable level. If the sentiment model is intended for your company specifically, stakeholders at the company should have an opportunity to review the sentiment output and provide feedback.

When the final sentiment model is ready, the Medallia research team merges it into a Experience Cloud release. If your company wants to process historical records with the newly-released sentiment model, contact Medallia Support to initiates a text reprocessing process.

For a list of languages with sentiment availability, see Text Analytics supported languages. For more information about how to enable sentiment for your program, see Configuring sentiment.