Generative Annotation (NEW)

Generative Annotation uses Microsoft’s Azure OpenAI endpoint to generate AI suggested labels to accelerate taxonomy design and early phases of model training; and reduce time-to-value for all Communications Mining use cases.

It includes:

Generative Annotation features will be automatically enabled on datasets – you don't need to do anything to start using them.

Once a dataset is created, cluster suggestions will automatically be generated within a short period of time. If a taxonomy has been uploaded (highly recommended), Communications Mining will suggest both existing and new labels for clusters.

When a taxonomy is uploaded to a dataset, this will also automatically trigger an initial model to be trained with no training data, just using label names and descriptions – this may take a few minutes from when you've uploaded the taxonomy.

Prerequisite: ‘Review and Annotate’ permission/

Cluster Suggestions will appear on the top of each Cluster page (white shading with blue border). This can be one or multiple suggested labels for each cluster.

If you have Label sentiment analysis enabled, Cluster Suggestions will have either positive or negative sentiment (white shading with green or red border).

You can tell it’s an AI suggested label by the red sparkle icon next to the label name.

Model trainers should review each Cluster Suggestion and:

Cluster Suggestions can significantly speed up the first phase of the model training process by automatically generating suggested labels for each cluster.

It can also help with taxonomy design, if users are struggling to define the concepts they want to train.

Cluster Suggestions are generated based on the identified theme shared across the messages within a cluster.

The creation of clusters and generation of label suggestions is an automatic and completely unsupervised process with no human input required.

Label suggestions on clusters will be generated with or without a pre-defined taxonomy, but suggestions will be influenced and typically made more helpful by leveraging imported / existing labels.

Prerequisite 1: ‘Review and Annotate’ permission.

Prerequisite 2: Imported list of label names.

Optional but highly recommended: Imported list of label descriptions.

Once the initial model has automatically trained using label names and descriptions as it's training input, predictions will appear for many of the messages in the dataset.

These predictions work in the exact same way as they have done previously – they are just generated with no training data.

If you have Label sentiment analysis enabled, initial predictions will have either positive or negative sentiment (different shades of green / red based on its confidence level).

Assisted Annotating works in any training batch or mode but it’s most effective to use in ‘Shuffle’ and ‘Teach Label’ (follow the regular annotating steps in each training batch in Train or Explore).

Assisted Annotating can significantly speed up the second phase of the model training process by automatically generating predictions for each label with sufficient context, with no training examples required.

Initial predictions will be driven by the quality of the label names and natural language descriptions (i.e. vague names might lead to vague or minimal predictions). Detailed label descriptions can boost the initial model’s performance.

As you train your dataset further, the platform will use both the label names and descriptions and your pinned examples to generate relevant label predictions.

These will keep improving with more training and ultimately rely only on annotated training examples when enough have been provided.

Assisted Annotating still requires supervised learning by accepting / rejecting the predictions, but it accelerates the most time-consuming part of model training by providing better predictions with zero or very few pinned examples.