Category Best Practices
Category classification is a logical grouping of calls based on linguistic information (topics) and/or non-linguistic information about those calls.
Classification takes place after the interaction is analyzed. During the classification, events and metadata are used to assign one or more categories to interactions. Category classification enables SpeechMiner® to find interactions with specific characteristics and to perform statistical analysis on groups of interactions based on those characteristics.
Use the following best practice tips to fine-tune your category definition:
- Before creating a category, gather information about the business issue you want to investigate (for example, ask yourself what type of interactions do you want SpeechMiner to recognize?).
- Determine the type of interactions generated by the business issue by listening to existing interactions about the subject or similar subjects, reading e-mails that relate to them, and consult agents who interact with customers. For example, for a customer service topic, the set of relevant interactions might include only incoming interactions about help and assistance.
- Verify the possible reasons for the interactions (for example, a specific campaign, customer service, sales opportunities, and so on).
- Gather examples of the types of interactions in which you are interested, to determine the best way to structure the Program.
- Give the category a simple name that is concise and unambiguous. For example, if you want to categorize interactions in which agents verify a customer with DOB during the first 30 seconds of call, the name should be "Verify DOB (first 30 seconds of call)". If you want to categorize interactions that signify customers that are contemplating leaving the company, the name of the category should be "At Risk".
- Use expressions to set the conditions an interaction must meet to be assigned to the category.
For example:- To categorize interactions in which agents verify a customer with DOB during the first 30 seconds of call, add the following expression: "Found topic "Verify DOB" in the first 30 seconds of the call".
- To categorize calls that may be considered "At Risk", add the following expression: "Found topic “Escalation” OR Found topic “Lawyer” OR Found topic “Report to Media”.
- Create a hierarchy of categories in which parent and sub-categories correspond to business objectives.
- Remember that when a regular category is converted into a parent category, the logical expression that defined the category is deleted.
- The goal of validating the category is to ensure precision of >80% at the lowest confidence possible.
- Distribute the category to support members and keep track of their progress as shown in the following table. Adjust the expression where needed:
| Category Name | Expression | Member Name | Date | Total Number of Interactions | Validated Interactions | Notes | Progress |
|---|---|---|---|---|---|---|---|
| Agent Skills : Empathy | Found Topic ‘Empathy’ | Steve | 12/3/2014 | 80 | 20 | >80% precision | Completed |
- Validate the category once a month to ensure that the category continues to classify information that results in meaningful reports. If, after 3 months of regular validation, a category remains stable in its precision, it can be validated every 6 months instead.
- To validate a category:
- Read the category expression and verify that you understand it. Consult the Lead if you do not understand the expression or the definition.
- In SpeechMiner select Search > Content Browser.
- Click Details
to review and understand the expression. - Click the category name to upload the interactions associated with the specific category.
- Open and review 10-20 interactions per category and verify that the expression is correctly categorizing interactions.
- Report all your progress to the lead. If FP events or expressions are failing, take detailed notes about why the category is failing. It may be necessary to adjust a phrase’s strictness level, delete a phrase, or audit more events to determine what the best course of action will be.
The number of interactions to be validated should be calculated based on the total number of interactions per category. For example, validating 5 interactions out of 300 interactions would not suffice to determine the precision of that category. For example, if the category expression is "Found topic "Dissatisfaction" at least once with Very-Low strictness", the interactions should include at least one "Dissatisfaction" TP event.