Text Analytics FAQ

What is the minimum volume required for implementing Text Analytics?

Medallia recommends a minimum, annual volume of 10,000 comments.

Can text analytics process non-survey data?

Yes. Text Analytics can process multiple non-survey data sources. For more information, see Import data for Text Analytics.

For companies with the social feedback offering, Medallia Text Analytics can be applied to social feedback data in addition to survey data. For information about including social feedback data from a configured social feedback system, see Configuring text analytics for social media data.

Can Text Analytics handle multiple surveys?

Yes. If necessary, each survey program can have its own topic set (which involves creating separate tag pools). For more information, see Creating a tag pool.

Can Text Analytics handle multiple comment fields per survey?

Yes. Companies can choose which comment fields to parse and translate. They can filter results by comment field as well.

How do we determine topics? Does Text Analytics suggest any?

You can create your own topics, or choose from a list of automated topics generated by Text Analytics using your data. To get ideas for topics, you can read through comments, and create topics about the subjects that are frequently discussed. For more information, see Build topics.

How does sentiment work?

Once a comment is matched with a topic, the sentiment algorithm looks at the portion (sentence or phrase) of the comment that discusses the topic. The sentiment algorithm then determines if that portion of the comment is strongly positive, positive, mixed opinion, negative, strongly negative, or has no opinion. When configuring sentiment for reports, you can decide whether to display the entire sentiment scale, or show positive and negative sentiment only.

For more information about sentiment, see Configuring sentiment.

How does Medallia create sentiment models?

The process of creating a new sentiment model always begins with a minimum of 40,000 sentences. Each of those sentences is labeled with a sentiment by a human. 40,000 is the minimum number of annotated sentences required to achieve generalization (to learn patterns inside the text). Using those initial sentences, the Medallia research team follows this general process:

  1. Use state-of-the-art machine learning techniques to discover features and patterns inside the text of annotated sentences and find boundaries that separate different sentiment classes (such as positive vs. negative) through an optimization process.
  2. After multiple iterations through the optimization process, select a set of features that optimize overall sentiment accuracy.
  3. After building the initial model, conduct tests to evaluate model accuracy (how accurately the model identifies the sentiment of new sentences). The goal is to achieve at least 80% overall accuracy across all segments of new data. Since the model tries to generalize, it's possible when analyzing a small data set accuracy might drop lower than 80%. Aside from conducting internal evaluations, the team engages companies to make sure they are satisfied before releasing the model.
    Note: Over time, the model can become less accurate if a large number of incoming sentences have not been learned by the model. The team continually engages with companies to maintain model accuracy and help companies gain insights from Text Analytics.

Can the parser process comments in multiple languages?

Yes. However, you must build topics in the languages that are parsed. For more information, see Configure language processing.

Is machine translation supported?

Yes. Comments can be parsed into one target language. For more information, see Translating comments.

How long does a regular implementation take?

A typical 50-topic implementation takes a minimum of 12 weeks.

How is the impact score calculated?

The impact score is calculated for each topic and can informally be described as your current customer satisfaction score, minus what your score would be if you excluded all surveys tagged by a certain topic. For a more detailed explanation, see Impact Score.