AI Explainer

The AI Explainer tool helps you understand how each word in a message contributes to triggering a specific intent. It provides insights into the underlying AI model and its decision-making process.

Overview

Prediction Probabilities

ViaSay Chatbot's AI does not return a single intent; instead, it returns a distribution of probabilities for each possible intent. The Explainer displays the top 5 intents with their prediction probabilities from this distribution.

Focus on Top 2 Intents

The Explainer focuses on the two most probable intents using two vertical bar charts. These charts show how much each word in the sentence contributes to the probability given to the intent.

For example, the word "Flights" might have a strong impact on the intent Book a Flight but not much on the intent Flight Status. This discrepancy might indicate a mistake in the training of the Flight Status intent.

One word can significantly impact different intents. For instance, the word "pizza" could strongly affect the intent Meal Onboard but have no impact on Flight Information.

Triggered Intent Overview

The Explainer provides an overview of how each word contributes to the triggered intent for a given message.

Current Limitations

The Explainer has some limitations and does not account for the following rules:

• Negative examples and confidence thresholds
• Rule-based intents
• Priorities
• Main user node vs. current user node rules
• Required entities

However, these rules are always considered in the standard NLP path within the widget.

References

Our Explainer tool utilizes an open-source library called LIME (Local Interpretable Model-agnostic Explanations). We estimate weights using local linear interpolation. We assume that for similar sentences, ViaSay Chatbot's AI model adds each word's weight, and we compute those weights accordingly.

💭

NLP models run independently between Draft and Production modes. This can lead to different results, so ensure you use the Explainer in the correct mode.