Training phrases

Creating an effective NLU for bots requires consideration of several factors. Here are a few practical tips for creating training phrases:

1. When preparing or adding new phrases, it's advisable to take a global point of view of the entire set of NLU. We should not focus solely on one process. It's valuable, whenever possible, to consider the entire NLU set and the bot's flow, and how a given phrase can impact its operation.
2. The greatest value while creating phrases lies in focusing on the most popular and likely queries. Adding edge cases to the NLU set typically results in decreased effectiveness and is ultimately not cost-effective. For processes already supported by the bot, we recommend that any added phrase generates at least 0.005% of traffic in production data.

Before making changes to NLU, it's advisable to conduct a brief analysis focusing on several key aspects:

1. Whether the phrase we want to add to NLU or cover in the flow answers the question. Avoid mapping phrases that do not answer the user's question.
   Question: Do you want to report the sale of a vehicle?
   Example of a phrase that does not answer the question: just left.
2. Whether the phrase will be 'understandable' for the bot. Phrases prepared for the bot's NLU should resemble communication between a human and a bot, not between humans. Communication should be concise, slogan-like, and clear—customers who understand that adopting such conversational style significantly increase their chances of understanding the bot.
   Question: Provide the number of travelers.
   Example of a phrase that would be unclear for the bot (and often for humans as well): whole family is coming.
3. Whether the phrase is unambiguous. Pay attention to the level of abstraction in statements and their context-dependence. Some phrases can be ambiguous in a given context. Please ensure that responses always match the questions asked.
   Question: What is the scope of your insurance?
   Example of an ambiguous phrase that should not be suggested as a leading phrase for any of the paths: "I don't know"
4. Is the phrase too descriptive? NLU should be constructed in such a way that the percentage of descriptive phrases compared to concise, specific phrases is minimized. Let’s try not to formulate phrases descriptively but directly.
   Question: Do you want to cancel your liability insurance at the end of the insurance period?
   Example of a phrase that is too descriptive: A friend recommended another company, and we want to move elsewhere.
5. Are the phrases leading to different paths not too similar to each other?