Phrases mapping

When creating NLU, it is important to properly group phrases. By having a set of phrases, we look at their entirety and seek out key elements among them. Instead of teaching the bot whole phrases like “I want to buy a policy,” “I want to activate a service,” “I want to report a claim,” it is better to break them down into elements like “buy,” “activate,” “policy,” “service,” “report,” “claim.” This way, we reduce the risk of content overlap between intents that would occur if we added intents with whole phrases.

Therefore, it is crucial to properly map phrases, which means determining the intents or entities that should be detected for a given phrase to guide the bot on the correct path. With a ready NLU model and wanting to map new phrases for another process, we analyze whether the elements making up the phrase have already been included as entities/intents. If not, we create a new entity/intent; otherwise, we simply list the key detections as conditions for phrase recognition. We determine the most important recognitions for a given phrase or group of phrases. If, during analysis, we encounter phrases that deviate from the rest and are hard to fit anywhere, we should consider whether these phrases are truly important. If not, it is better to discard them rather than adjust the NLU set just to fit them in or create special elements solely for them. Additionally, with production data, we can check how often a given phrase is used in a specific context and include it only when its occurrence is indeed frequent. If a phrase is too general, ambiguous, or does not answer the asked question, it is better to implement an additional question in the process rather than include it as a guiding phrase.