Dialog Context for Language Understanding Accuracy

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing dialog systems face challenges in accurately determining the intent of short and ambiguous user utterances due to their stateless nature, which limits their ability to leverage prior context for disambiguation.

Innovation Solution

The proposed solution involves utilizing prior dialog context to improve the language understanding model's ability to disambiguate intents and entities in user utterances by training the model with utterance-context pairs, allowing it to consider both the current utterance and previous intents and entities for more accurate interpretation.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If a stateless language understanding model is used, then the system is simpler and faster, but the accuracy of intent detection for short and ambiguous utterances deteriorates

Engineering Contradiction:
Improveintent detection accuracyVSAvoidmodel architecture complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent transforms the stateless model into a stateful model by adding a temporal dimension through dialog context. The language understanding model now processes not only the current utterance but also historical dialog context, effectively moving from a 1D (current input only) to a 2D (current input + historical context) processing space. This dimensional expansion enables the model to disambiguate short utterances by leveraging information from previous turns in the conversation.

Inventive Principle:
Principle #17Another dimension (Dimensionality change)

2Measurement precision

If prior dialog context is incorporated, then the accuracy of understanding short and ambiguous utterances improves, but the processing time and computational resources increase

Engineering Contradiction:
Improveutterance disambiguation accuracyVSAvoidprocessing time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The patent extracts only the relevant portions of dialog context (previous intents and entities) and feeds them selectively to the language understanding model. Instead of processing the entire dialog history, the system identifies and extracts key contextual elements that are most useful for disambiguation, thereby reducing the computational burden while maintaining improved accuracy.

Inventive Principle:
Principle #2Taking out (Extraction)

3Adaptability or versatility

If the language understanding model considers prior context, then the ability to interpret ambiguous utterances improves, but the system's stateless architecture becomes more complex

Engineering Contradiction:
Improvecontext adaptation capabilityVSAvoidsystem architecture complexity
Core Design Contradiction:
Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The patent performs preliminary processing of dialog context by pre-identifying and structuring previous intents and entities before they are needed for disambiguation. The context is prepared in advance and organized in a format that can be efficiently consumed by the language understanding model, reducing the complexity of real-time context integration while enhancing adaptability.

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP3963572B1Using dialog context to improve language understanding
Publication Date: 2025.02.05 MICROSOFT TECHNOLOGY LICENSING LLC
  • EP3963572B1 patent drawingFigure 1
  • EP3963572B1 patent drawingFigure 2
  • EP3963572B1 patent drawingFigure 3

AI summary

Described herein is a mechanism for improving the accuracy of a language model interpreting short input utterances. A language model operates in a stateless manner, only ascertaining the intents and/or entities associated with a presented input utterance. To increase the accuracy, two language understanding models are trained. One is trained using only input utterances. The second is trained using input utterance-prior dialog context pairs. The prior dialog context is previous intents and/or entities already determined from the utterances in prior turns of the dialog. When input is received, the language understanding model decides whether the input comprises only an utterance or an utterance and prior dialog context. The appropriate trained machine learning model is selected and the intents and/or entities associated with the input determined by the selected machine learning model.