分类设计最佳实践

关键分类元素

• Number of labels: Typical datasets have 50-100 labels, but this number can vary depending on the objectives for a dataset. An effective use case can have much fewer than 50 labels. The system imposes a limit of 200 labels for a dataset because beyond this point, the taxonomy becomes very difficult to manage and train, and it leads to reduced performance.

• Label names: Label names should be concise and descriptive because the Generative Annotation feature uses them as a training input to speed up and improve the training process. You can always edit them, but to ensure they display effectively in the platform UI, a character limit of 64 characters is set for any given label, including its levels of hierarchy.

• Label descriptions: Add natural language descriptions to your labels because they are used as a training input by the Generative Annotation feature for automatic training. Descriptions also help ensure annotating consistency among model trainers and provide helpful context to others viewing the dataset for analytical purposes.

构建分类结构

我们建议您遵循以下最佳实践，以正确构建分类结构并确保模型具有高性能：

• Align with objectives: Make sure each label serves a specific business purpose and is aligned to your defined objectives. If your dataset is meant for automation, many labels should match the specific requests needed for downstream processing. If your dataset is meant for analytics (or both), include additional labels that cover concepts like issue types, root causes, and quality of service issues such as chaser messages, escalations, and disputes.

• Be distinct: Each label should be specific and not overlap with other labels.

• Be specific: Avoid broad, vague, or confusing concepts, as they are more likely to perform poorly and provide less business value. Split broad labels into multiple distinct labels if possible. Start with specific labels, such as more levels of hierarchy, and merge them later if needed, rather than breaking down broad labels manually.

• Be identifiable: Ensure each label is clearly identifiable from the text of the messages it is applied to.

• Use parent labels: If you expect to have many similar concepts related to a broader topic, use a parent label.

• Use child labels:Make sure that every label nested under another label is a subset of that label.

• Limit hierarchy levels: Try not to add more than four levels of hierarchy, as the model becomes increasingly complex to train.

• Include uninformative labls: Create some non-value-adding labels, such as thank-you emails, so you can tell the platform what is or isn’t important to analyze.