Attention-Based NLU Classification With Built-In Slot Explainability
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing artificial intelligence models for natural language understanding lack explainability, making it difficult to improve the model when errors occur and reducing user trust due to unclear decision-making processes.
Innovation Solution
A method is introduced to provide both classification and explainability by processing utterances through intent classification, generating slot type weights and binary classifier logits, and visualizing the process using a graphical slot explanation, which includes parsing utterances into vectors, encoding with an encoder, applying intent and slot classifiers, and performing feature fusion to output slot logits in a two-column format.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of information
If an additional process is added to provide explainability, then explainability is improved, but efficiency and correctness of the model are reduced
Solution Approach 1:
The patent merges the explainability generation process with the intent detection and slot filling processes by using shared neural network components and joint training. The visualization network uses the same encoded representation and attention mechanisms as the classification tasks, eliminating the need for separate post-processing and improving both efficiency and correctness.
Solution Approach 2:
The neural network model is designed to perform multiple functions simultaneously: intent classification, slot filling, and explainability generation. The same encoder and attention mechanisms serve all three purposes, making the system multi-functional and avoiding the overhead of separate dedicated processes.
2Loss of information
If an additional process is added to provide explainability, then explainability is improved, but correctness of the model is reduced
Solution Approach 1:
By combining the explainability generation with the classification tasks in a unified neural network framework, the patent ensures that the explanations are generated from the same learned representations that drive accurate classification. This joint learning approach ensures correctness while providing explainability.
Solution Approach 2:
The model uses self-attention mechanisms that provide feedback loops where the encoded representation is repeatedly refined by attending to different parts of the input. This iterative refinement ensures that the final representations used for both classification and explanation are highly accurate and reliable.
3Productivity
If features are learned collectively over all slot types, then model efficiency is improved, but explainability is lost
Solution Approach 1:
The patent segments the slot filling task into individual slot-type-specific classification heads while maintaining a shared encoder. This allows the model to learn features collectively in the shared representation layer (maintaining efficiency) while providing slot-specific attention weights for explainability (restoring interpretability).
Solution Approach 2:
The model applies local quality by generating separate attention weights and explanations for each slot type based on the shared encoded representation. Each slot gets its own customized explanation tailored to its specific classification needs, while benefiting from the collective feature learning of the shared encoder.
Data Source
AI summary
In an artificial intelligence model (AI model), input data is processed to provide both classification of the input data and a visualization of the process of the AI model. This is done by performing intent and slot classification of the input data, generating weights and binary classifier logits, performing feature fusion and classification. A graphical explanation is then output as a visualization along with logits.