Tips for building your natural language understanding (NLU)

To achieve optimal performance and reliability of your NLU system, follow these best practices when designing and testing intent classification.

Maintain balanced training data across intents

Disproportionate numbers of training examples per intent can introduce bias into your model. If one intent has a considerably larger amount of training examples than others, the model may favor that intent when uncertain.

Ensure a relatively even distribution of training phrases across all intents. If you intentionally want to prioritize an intent, do so consciously and document the reasoning.

Understand word vector similarity

NLU systems use vector representations (embeddings) to understand word meanings based on context. Words with similar meanings or contexts, like cash back and refund, may be mapped closely in the model’s vector space. This can lead to misclassification if you do not clearly distinguish the intents they belong by context.

Use diverse, representative training examples for each intent, emphasizing unique contextual clues that differentiate one intent from another.

NLU is a classification problem

NLU models classify input text by assigning probabilities to each intent. For N intents, the model returns a probability distribution over all N intents. The intent with the highest probability is selected, provided it meets the confidence threshold.

Review full probability distributions when troubleshooting misclassifications, not just the top-scoring intent.

Understand confidence thresholds and intent fallbacks

The system typically requires a minimum confidence score to classify an input into a specific intent. The NLU model requires at least 40% (0.4) confidence to assign an intent. If the top intent falls below this threshold, the system returns a fallback or None intent.

Leave the threshold at the default value unless you have a specific reason to adjust it. Raising it too high may lead to excessive fallback responses; lowering it may increase incorrect intent matches.

Tips for generating intent and utterance with AI

Define clear, non-overlapping intents

To avoid ambiguity, ensure that each intent serves a distinct purpose. Overlapping intent definitions lead to confusion during training and lower classification accuracy.

Establish clear boundaries between intents. If two intents are too similar, consider merging them or refining their definitions.

Start with a core set of intents

Begin with a small, high priority set of intents before you expand. A focused set helps you iterate faster and build a solid foundation.

Develop a proof-of-concept model with essential intents. Expand only after the core experience is works reliably.

Use consistent naming patterns

Standardize the way that you name intents across your project. Consistency improves maintainability and helps teams understand intent purposes immediately.

Tips for testing your NLU

Avoid one-word inputs during testing

One-word inputs are typically too ambiguous and lack the necessary context for accurate intent classification. NLU models rely on patterns and context – without surrounding words, the model must guess based on limited information and may default to the most common or closely related intent.

Use natural, full-sentence queries that reflect real user behavior.

Prepare high-quality training data

Use these tips:

• Develop a viable training set: 10–20 diverse examples per intent to start.
• Vary phrasing: Include synonyms, different syntactic structures, and common typos.
• Avoid overlap: To reduce confusion, make sure utterances clearly distinguish between intents.
• Test coverage: Sample utterances from all intents manually to check NLU predictions.

Tips for using the intent miner

Use Intent Miner to build intents and utterances tailored to your customers way of phrasing for better results in intent detection.

Intent miner searches a historical set of conversation chat and voice transcripts between agents and customers for a specified date range. Next, it extracts the top set of intents and the analytics that surround those intents. Finally, it returns a list of intent utterances that the system associates with that intent.

You can import mined intents in Architect into a Genesys Dialog Engine Bot Flow or Genesys Digital Bot Flow for intent configuration. For more information, see Import mined intents into a Genesys Dialog Engine Bot Flow or Genesys Digital Bot Flow.