Dynamic Memory Network Dialog State Tracking

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Conventional dialog state tracking systems are inflexible, inaccurate, and inefficient due to their inability to retain and consider pertinent information from previous dialog segments, leading to inaccurate and unhelpful responses that require additional user interactions and increased computing resources.

Innovation Solution

A neural network with a dynamic memory network architecture that uses multiple memory slots and gating mechanisms, such as reset and update gates, to modify memory slot values based on cross-slot interactions, allowing for flexible and accurate generation of digital dialog states that consider all relevant information.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Speed

If conventional dialog state tracking models continuously update datastores with new dialog segment data, then the system processes new information efficiently, but the system pushes out (forgets) old previously stored data even when still relevant

Engineering Contradiction:

Improveprocessing speed of new dialog segmentsVSAvoidloss of pertinent information from previous dialog segments

Core Design Contradiction:

SpeedVSLoss of information

Solution Approach 1:

The patent implements a dynamic memory network where memory slots and their associated gates (reset gate, update gate, read gate, write gate) can dynamically adjust their states based on the relevance of information. The memory network transitions between different operational modes (reading vs. writing) depending on whether the current dialog segment contains new information or refers to previously stored information, allowing the system to adaptively retain or update information as needed

Inventive Principle:

Principle #15Dynamics

Solution Approach 2:

The patent changes the parameter of information retention by introducing gating mechanisms that control the flow of information in and out of memory slots. The reset gate and update gate modulate the memory cell values to either preserve old information or incorporate new information based on the dialog context, while the read gate and write gate control information access and storage operations

Inventive Principle:

Principle #35Parameter changes

2Productivity

If conventional dialog state tracking systems rigidly determine current dialog state without considering all relevant previous segments, then the system operates efficiently with simple processing, but the system generates inaccurate dialog state predictions

Engineering Contradiction:

Improveoperational efficiency of dialog state trackingVSAvoidaccuracy of dialog state determination

Core Design Contradiction:

ProductivityVSMeasurement precision

Solution Approach 1:

The patent implements feedback mechanisms through the dynamic memory network where the system continuously reads from and writes to memory slots based on dialog segment analysis. The gating mechanisms provide feedback control by adjusting memory states based on the relevance of current and previous dialog segments, allowing the system to refine its dialog state predictions iteratively while maintaining operational efficiency

Inventive Principle:

Principle #23Feedback

Solution Approach 2:

The patent performs preliminary actions by pre-processing dialog segments and storing extracted information in memory slots before final dialog state determination. The system prepares memory representations of key dialog elements in advance, allowing for faster and more accurate state prediction when needed

Inventive Principle:

Principle #10Preliminary action

3Reliability

If conventional systems require additional user interactions to arrive at correct responses, then the system can ensure accuracy through multiple interactions, but the system increases computing resource consumption and reduces user experience

Engineering Contradiction:

Improveaccuracy of response generationVSAvoidcomputing resource consumption

Core Design Contradiction:

ReliabilityVSUse of energy by moving object

Solution Approach 1:

The patent replaces the mechanical interaction loop (user asks question → system responds → user clarifies → system responds) with a cognitive simulation system that uses dynamic memory networks to anticipate and resolve ambiguities internally. The system simulates human-like memory and reasoning processes to determine dialog state and generate responses without requiring actual back-and-forth interactions, thereby reducing computing resource consumption while maintaining response accuracy

Inventive Principle:

Principle #28Mechanics substitution (Replace mechanical system)

Data Source

PatentUS11657802B2Utilizing a dynamic memory network for state tracking

Publication Date: 2023.05.23 ADOBE INC

AI summary

The present disclosure relates to generating digital responses based on digital dialog states generated by a neural network having a dynamic memory network architecture. For example, in one or more embodiments, the disclosed system provides a digital dialog having one or more segments to a dialog state tracking neural network having a dynamic memory network architecture that includes a set of multiple memory slots. In some embodiments, the dialog state tracking neural network further includes update gates and reset gates used in modifying the values stored in the memory slots. For instance, the disclosed system can utilize cross-slot interaction update/reset gates to accurately generate a digital dialog state for each of the segments of digital dialog. Subsequently, the system generates a digital response for each segment of digital dialog based on the digital dialog state.