What are Metadata Correlations used for? 

Metadata Correlations are identified automatically by the system if there's a statistically significant relationship between comments and the accompanying metadata.  

Sample Use Cases

Segmenting customers: Comments & Demographic data

Find out if different categories of customers value different things using demographic metadata: For example, male vs. female, different age groups, level of education. 

Prioritizing customers: Comments & Satisfaction rating

Do you need to focus on making happy customers even happier or preventing the unhappy ones from churning? Prioritize where to take action based on the segment you're interested in. 

Ensure consistent service across all locations: Comments & Location data

Are certain locations doing a lot better or a lot worse? Learn from high-achieving locations and improve the ones lagging behind. 

We will explore a similar example below.


How do I use Metadata Correlations?

Here's a step-by-step walkthrough. In this example we are using a dataset comprised of international homestay reviews.

1. Upload the necessary data

Along with unstructured data such as comments, you'll need to upload accompanying metadata such as location, satisfaction rating, demographic data or structured data specific to the dataset such as room type. Here is an example:

Column C contains the Review Content. These are the comments that will be analyzed with Keatext’s NLP engine. Columns A, B, D, E, F and G contain structured metadata. Once the analysis is complete, the system will calculate and highlight any interesting relationships.

2. Let the analysis finish

The analysis needs to be complete before the system compares the distributions. It does this by automatically running a chi square test,  and then displays the statistically significant correlations. Attempting to see the correlations before the analysis is done will result in an error message. 

3. Investigate the highlighted Metadata Correlations

For each identified Comment, the Metadata Correlation is computed individually. In this example, we are interested in the Comment “place was very clean.'' Click on the row to reveal the analytics summary for that specific Issue. 

Underneath the summary panel, click on the Metadata Correlation tab to reveal what the system found automatically. 

Metadata Correlations are displayed in two ways, a donut chart and a table. The latter is helpful when there are many values. 

In the example, 63% of people who said "place was very clean" (or variations of the same idea) rented a homestay in the United States. 

4. Explore other relationships

It is also possible to add correlations manually, but it’s up to the you to decide if they are significant or not. Scrolling below the correlation chart or table will reveal the Create Correlation button. 

To add a correlation, you needs to select the data Source and the relevant Field (column from the dataset).  

How accurate are Metadata Correlations?

The purpose of the Metadata Correlations is to highlight interesting information. Statistical relevance depends on the amount of data uploaded. We recommend users upload several thousand responses to reach statistically significant insights. 

Did this answer your question?