Learning Relationships Between Projects (Legacy AI vs New AI)

Learning relationships allow you to reuse topics from one project in another.
This can save time and help maintain consistency across similar datasets.

If you’re looking for instructions on how to set up learning relationships, please refer to the following article.

Below you’ll find a clear explanation of how learning relationships behave in both the Legacy AI and the New AI, and when you should (or shouldn’t) use them.

1. Learning Relationships in the Legacy AI

In the Legacy AI, learning relationships play an important role in ensuring consistency across multiple related projects or waves.

A learning relationship allows Project B to inherit both:

The topic structure (topics, categories, sentiment), and
The training you performed in the Project A.

This means the AI will classify new comments in Project B using the patterns structure learned from Project A.

Learning Cascades (Multiple Projects)

If you connect several projects, e.g., Project A → Project B → Project C, the AI forms a learning cascade:

C learns from B
B learns from A
A indirectly influences C

However, as the chain grows very long, the influence of the earlier projects becomes weaker.
For example:

With 3–5 connected projects → still strong
With 10–20 projects → the earliest projects have minimal impact

Best practice for Legacy AI:

Choose the most recent, best-trained project as the learning source whenever possible.

What happens if you remove the learning relationship?

Immediately: Nothing changes. Your existing topics and assignments remain intact.
Over time: Once you start adjusting topics or making assignment changes, the AI will no longer reinforce the inherited learnings. This can cause the model to gradually “forget” what it inherited from the previous projects.

Ensuring New Topics Aren’t Missed

When using learning relationships, especially with multiple project waves, consider:

a) Generating Additional Topics

After selecting the learning project, use the “Additional Topics” feature to detect emerging patterns specific to the new dataset.

b) Manual Review

Review a sample of comments without topics or use free-text search to catch new themes that may not exist in the inherited project.

2. Learning Relationships in the New AI

The behavior of learning relationships changes significantly in the New AI.

⭐ Important: For the New AI, we generally do not recommend using learning relationships.

Here is why:

1. There is a risk of confusing the New AI when inheriting from another project

Inheriting topics from another project can actually hurt performance if:

That project wasn’t trained well, or
It was trained using the Legacy AI, which works differently.

In these cases, the inherited topics may not match your new dataset and can confuse the New AI, leading to lower-quality topic assignments.

2. Inherited topic descriptions often do not fit the new dataset

Internal testing shows that if the new project contains:

Even slight variations of the same question, or
Different data column context,

The inherited descriptions may no longer match well.
This often results in:

Lower topic assignment confidence
More unassigned rows
Inconsistent labeling behavior

3. Adding new topics can increase unassigned rows

When you inherit topics and then add new ones, the AI tries to interpret:

Old descriptions intended for the previous dataset
New descriptions written for your current dataset

This mismatch can further reduce assignment quality.