Improving the Quality of Your Text Analysis Classifications – Displayr Help

After you've gotten started with text analysis, there are ways that you can tidy up and improve your classifications. Read on for the next steps to refine your initial text analysis work.

Resolving Anything Left Unclassified
Checking the Quality of Classified Responses
Troubleshooting Classifications
Editing Your Classifications

Resolving Anything Left Unclassified

As with all text classification, it's always a good idea to review the work done and clean up anything outstanding. Once the AI has run, the screen will look something like this:

Notice the "Unclassified" theme on the left; some responses remain unclassified. In the example, there is a text response "undefined" without an associated theme and is left unclassified. "Undefined" doesn't represent anything meaningful, so let's classify it as "Missing Data".

Select the response "undefined" and drag and drop it onto the "Missing Data" theme, or hover over "undefined", click Manually classify, start typing "Missing Data", and select that theme to assign the response.

Next, review any remaining unclassified responses and classify them. You can create a new theme, such as "All other responses" to capture the responses that don't fit into existing themes.

Set the Theme filter dropdown in the Responses section to Unclassified to show all unclassified responses.
Select all of the unclassified responses. You can hold down the shift key while selecting the first and last text responses to select multiple items.
Drag the responses below the existing Themes and drop them into the purple box that appears.
Give the new theme a label, e.g., "All other responses" and hit Enter.

Checking the Quality of Classified Responses

It's important to give any automatic classification a quality check before reporting on your analysis. A quick way of doing that is to select a theme from the Theme filter and quickly scan through for a quality check.

If there are more than a few items that you need to reclassify into an entirely different theme, drag them all to the "Unclassified" theme. Then click Classify again to re-run the AI theme assignments. There are a number of ways that you can improve how text is classified; see Troubleshooting Classifications below.

Below, I've selected the "Brand loyalty" theme and can see that there are some items, such as "drink for the enjoyment and the fun it brings", that are better placed in another theme.

You can manually reclassify responses:

Select the text response.
To add it to an additional theme, drag and drop it into the new theme, or start typing the new theme's label and select it.
To remove it from the current assigned theme and add it to another, click on the triple letters in the Classification column, remove the assigned theme by clicking on the X to the right of the label, start typing the correct theme, and hit Enter.

Below, "drink for the enjoyment and the fun it brings" is added to "Social perception" and removed from "Brand loyalty".

You can delete existing themes and reclassify them into an existing theme by right-clicking and selecting Delete and Reclassify as (and then selecting the theme to classify into), or delete a theme altogether by right-clicking and selecting Delete (Del).

Alternatively, you can add more themes manually or run the Create function again to create additional themes using AI.

See How to Refine and Edit Text Themes After Categorization for more details and examples for refining and editing your text themes.

Troubleshooting Classifications

You may find that the AI has struggled to classify a chunk of responses, or that some responses have been poorly classified when you do your quality check.

To improve the overall quality, consider the following tricks:

Edit the themes to make them more explicit by adding or editing the label to make sure that there's more overlap between the theme label and the data. Right-click an existing theme and select Rename to add to or edit the label.
If there is lots of variability in the data, e.g., if the responses are long-format and people say a lot of different things, consider adding more themes either manually (via + Add New Theme) or by using the Create function for the AI to scan through unclassified responses.
You can help the AI improve by manually classifying some responses to existing themes so that it has some examples to learn from.
Customize the prompt when classifying responses.

Editing Your Classifications

If you decide you want to revisit your classifications later on, either to review or update them, or if you've updated your input data, you can edit your existing classifications. Click the variable with "Categorized" in the name in the Data Sources pane, and then from Properties , click Data > Attributes > Text Categorization.