Theme Structure review: Best Practice 🎓

This article contains best practice guidance and the answer to FAQs about our Theme Structure and how to review it!

Written by Mikhail Dubov
Updated over 3 weeks ago

Theme Structure Review

What is a theme structure?

Your theme structure is your unique neural network! 🧠

When Chattermill first analyses your data, our Machine Learning team creates a “theme structure” by examining what topics are consistently being mentioned. These topics, known as themes, are then organised into categories that suit your business.

This is what will allow your model to tag all of your comments with theme mentions, each of which will have a negative, positive or neutral sentiment.

How should I review my theme structure?

1. Please directly @ mention in a comment your ML Ops Analysts with any feedback/change requests.
2. We recommend sharing the theme structure only with those stakeholders who will also be responsible for reviewing tagging (a smaller group ensures alignment and version control).
3. Our Machine Learning team are experts in creating theme structures that are representative of your data and optimised for model performance and accuracy. As you review your theme structure, please keep in mind that certain change requests may increase the risk of impacting model performance.

FAQs

Can I add themes which aren't on the spreadsheet?

• Yes. As long as there is a sufficient number of mentions, we can add a new theme. A sufficient number of mentions is required per theme to ensure model accuracy when tagging.
• For any additional theme suggestions, please propose a theme name, a definition/any associated keywords and provide an example comment if possible.

If you would like a topic which does not yet have enough mentions to be its own theme, you can use custom tags or segments to track it.

Can I merge themes together?

• We create a theme when there is a high volume of mentions for it and/or when its concept is unique compared to other themes. However, if your organisation considers two or more themes to be conceptually similar we can consider merging them as long as neither one of them has a large number of mentions.

Can I split a theme into two or more?

• A single theme can be split into two or more if the mentions and concepts are distinct enough to be separated out (and there is sufficient volume per theme). As a rule of thumb, if a person without in-depth knowledge of your organisation was able to distinguish between these themes, then there is a good chance that our algorithm will be able to do so!

Can I delete/remove themes?

• The themes proposed to you were all detected as being of significant conceptual importance and/or frequency. We are unable to remove any themes with a high volume of mentions and which represent a significant proportion of your data. Removing themes would mean attributing them to a “Miscellaneous” or other broad category, which risks over indexing within your algorithm and compromises accuracy.

Can I rename categories or themes?

• Yes. You can rename any theme or category to better suit your business needs. Simply confirm with your CSM the current theme/category name and what you’d like to change it to. Please note no two themes or categories can have the same name.

Can I change the definition of themes?

• Yes. However, to change the definition of a theme you will need to share with our team as much context as possible. This will allow us to successfully train the theme according to your preferences.

Can I move themes to different categories?

• Yes. Themes can be moved into different categories. At least one theme is required per category and we recommend keeping relevant/related themes together within the same category.

Can themes have sentiment in the name or definition, i.e. 'Happy with Product' or 'Unhappy with Product'?

• No, sentiment cannot be accounted for in theme names or definitions as we operate on sentiment neutrality. Sentiment will be assigned for every theme based on comment content.