Before designing your taxonomy, it’s important to understand what should be captured by labels, entities, and metadata, to meet your objectives. There should be minimal overlaps as they all complement each other.
- Concepts, themes and intents
- E.g. ‘Change of address request’, ‘Urgent’, ‘Status update request’, etc.
- Should not be used to capture information that is present in the metadata
- Structured data points extracted from the text
- E.g. Policy numbers, trade IDs, URLs, dates, monetary quantities, etc.
- Additional structured information associated with each message
- Metadata properties can be user properties (defined and added pre-upload, e.g. NPS score), email properties (captured from emails, e.g. sender, recipients, domains, etc.), and thread properties (automatically derived by Re:infer for threaded data like emails and chats, e.g. # of messages in thread, thread duration, etc.)
Here are some of the key distinctions and similarities between labels and entities. The two are typically used in combination for automation, but individually they serve different purposes:
What can Re:infer learn from during training?
Re:infer makes label predictions based on text of the verbatim (for emails, this means the subject and body of the email), as well as some metadata properties. For entities, it learns from the assigned span of text, and the context of the text surrounding that span.
Using labels, entities and metadata together
Below is an example verbatim that shows how labels, entities and metadata are distinct, but complementary to one another. For this inbound request to be automated, each of them may be required for a specific purpose: