Guided Dialogue Generation with Selective External Search
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing language generation neural networks face inefficiencies in providing accurate responses without excessive communication bandwidth usage or computational overhead, particularly when relying solely on external searches or retraining.
Innovation Solution
A hybrid solution combining a trained language generation neural network with a response selection neural network and rule violation detection network to facilitate limited searching and rule-based filtering, allowing for efficient and accurate information provision.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a language generation neural network relies solely on external searches to provide accurate responses, then response accuracy is improved, but communication bandwidth usage increases
Solution Approach 1:
The system performs preliminary actions by pre-training the language generation neural network with extensive knowledge during offline training phases. This allows the model to have internalized knowledge ready for immediate use without requiring online searches, reducing communication bandwidth while maintaining response accuracy for common queries.
Solution Approach 2:
The patent introduces an intermediary mechanism that determines when external searches are actually necessary. The system uses a hybrid approach where the neural network first attempts to generate responses from its internal knowledge, and only performs external searches when the uncertainty or confidence threshold indicates the need for additional information verification.
2Reliability
If a language generation neural network performs extensive external searches to ensure accurate responses, then response reliability is improved, but computational overhead increases
Solution Approach 1:
The system applies partial action by performing external searches selectively rather than extensively for every query. The hybrid architecture enables the model to perform just enough searching to achieve reliable responses without unnecessary computational overhead, using confidence thresholds to determine when partial searching suffices.
Solution Approach 2:
The patent implements local quality by applying different processing strategies to different parts of the knowledge space. The system uses localized external searches only for specific topics or queries where the neural network's internal knowledge is insufficient or uncertain, rather than uniformly searching for all queries, thus reducing overall computational overhead while maintaining reliability.
3Reliability
If a language generation neural network is retrained frequently to provide updated information, then information currency is improved, but computational inefficiency increases
Solution Approach 1:
The system performs preliminary knowledge updates during offline training phases rather than frequent online retraining. The neural network is periodically retrained with updated data in advance, allowing the model to maintain current information without incurring the computational costs of frequent retraining cycles during deployment.
Solution Approach 2:
The patent introduces an intermediary external search system that acts as a mediator between the need for current information and the cost of retraining. Instead of retraining the model frequently, the system uses targeted external searches to retrieve updated information when needed, combining the benefits of current information with computational efficiency.
4Loss of energy
If a language generation neural network performs no external searching to reduce bandwidth usage, then communication efficiency is improved, but response accuracy deteriorates
Solution Approach 1:
The system performs preliminary knowledge acquisition during the offline training phase, pre-loading extensive knowledge into the neural network's parameters. This allows the model to operate with minimal external searching during deployment, reducing communication bandwidth while maintaining response accuracy for the majority of queries that fall within its trained knowledge domain.
Data Source
AI summary
Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for enabling a user to conduct a dialogue. Implementations of the system learn when to rely on supporting evidence, obtained from an external search system via a search system interface, and are also able to generate replies for the user that align with the preferences of a previously trained response selection neural network. Implementations of the system can also use a previously trained rule violation detection neural network to generate replies that take account of previously learnt rules.