Hierarchical Routing Algorithm Generation for Virtual Speech Assistants
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing distribution algorithms for virtual voice assistants face increasing complexity and inefficiency due to the growing number of specific models and similar sentences, leading to longer query training and execution times and poor runtime performance.
Innovation Solution
A method involving detection of changes in specific models, optimization of input data, and application of subsampling techniques to generate refined distribution algorithms, including domain models and global dispatchers, which are validated and implemented in virtual voice assistants.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If the number of specific models and sentences is increased to expand virtual assistant functionality, then the coverage and capability of the virtual assistant is improved, but the complexity of the distribution algorithm increases and runtime performance deteriorates
Solution Approach 1:
The distribution algorithm is segmented into multiple hierarchical levels: a global dispatcher at the top level and domain-specific dispatchers at lower levels. This segmentation allows the system to handle diverse functionalities while maintaining manageable complexity at each level, as each dispatcher only needs to handle a subset of the total functionality.
Solution Approach 2:
The patent introduces a hierarchical dimension to the distribution algorithm structure. Instead of a flat single-level dispatcher, the system organizes dispatchers in a tree-like hierarchy where the global dispatcher routes to domain dispatchers, which in turn route to specific models. This dimensional organization resolves the complexity issue by distributing the routing decision-making across multiple levels.
2Adaptability or versatility
If more sentences are added to specific models to improve functionality, then the virtual assistant can handle more user requests, but the training and execution time of the distribution algorithm increases
Solution Approach 1:
The training process is segmented into multiple stages corresponding to the hierarchical levels. Each dispatcher is trained independently on its specific subset of data, rather than training a single monolithic dispatcher on all data. This segmentation reduces the computational burden and training time for each individual model while maintaining overall system functionality.
Solution Approach 2:
The global dispatcher is trained and established first to perform preliminary routing of queries to appropriate domains. This preliminary action filters out the majority of routing decisions at the top level, so that domain-specific dispatchers only need to handle a subset of queries, thereby reducing their training and execution time requirements.
3Measurement precision
If similar sentences are included in the distribution algorithm to improve accuracy, then the virtual assistant can distinguish between similar requests, but the algorithm size and processing complexity increase
Solution Approach 1:
Similar sentences are handled by segmenting them into different domain-specific dispatchers rather than requiring a single large dispatcher to handle all variations. Each domain dispatcher specializes in recognizing similar sentences within its specific domain context, reducing the overall algorithm size while maintaining distinction accuracy.
Solution Approach 2:
Each domain dispatcher is given specialized local quality or expertise for handling similar sentences within its specific domain. Instead of one dispatcher trying to learn all variations of similar sentences across all domains, each dispatcher develops specialized knowledge for its domain, improving accuracy while keeping individual dispatcher sizes manageable.
Data Source
Figure 1
Figure 2
AI summary
The present invention relates to techniques for generating a distribution algorithm for a virtual voice assistant using a computing unit, wherein the distribution algorithm is configured to forward a user request to a specific evaluation module for processing the request, wherein the generation of the distribution algorithm comprises the following steps: • generating an input data set, wherein the input data set comprises elements, in particular elements of the specific evaluation modules, especially the specific domains; • optimizing the input data set with a view to generating a reduced input data set; • applying at least two different methods to the generated reduced data set to generate at least two different subsets, in particular a first subset and a second subset, from the reduced data set;• Generation of at least two different distribution algorithm models, each of which is generated based on one of the subsets; in particular, a first distribution algorithm model is generated based on the first subset and a second distribution algorithm model is generated based on the first subset of the reduced dataset; • Testing the quality of the at least two different distribution algorithm models using reference data, assigning a quality score to each of the distribution algorithm models as a test result; • Selection of the distribution algorithm model with the best quality score as the distribution algorithm.