Unified Spoken Request Parsing for Multi-Intent Voice Assistants
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Intelligent voice assistant products face complex system models, high training and iteration costs, and poor cross-lingual migration capabilities due to the use of multiple separate models for different spoken language understanding scenarios, limiting their effectiveness in handling multi-intent and multi-round requests.
Innovation Solution
A unified prediction model is employed to predict sub-requests and their corresponding instruction information from spoken language requests, using a generative large language model trained with chain-of-thought to break down requests into sub-requests and determine intents, domains, and slots, thereby simplifying the system and improving understanding capabilities.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If multiple separate models are used for different spoken language understanding scenarios, then the system can handle various scenarios (multi-intent, single-intent), but the system model becomes complex and training/iteration costs increase
Solution Approach 1:
The patent merges multiple separate models (multi-intent model, single-intent model, multi-round model) into a unified model that can handle all spoken language understanding scenarios. The unified model receives spoken language requests and directly outputs sub-requests with their corresponding intents, domains, and slots, eliminating the need for multiple separate models and reducing system complexity while maintaining comprehensive scenario coverage.
Solution Approach 2:
The unified model is designed to perform multiple functions simultaneously - it can handle multi-intent requests, single-intent requests, and multi-round conversations within a single model architecture. This multi-functional design allows the model to adapt to different spoken language understanding scenarios without requiring separate specialized models, thereby reducing overall system complexity.
2Adaptability or versatility
If multiple separate models are used for different spoken language understanding scenarios, then various scenarios can be supported, but training and continuous iteration costs become high
Solution Approach 1:
By combining multiple scenario-specific models into a single unified model, the patent reduces the total computational resources required for training and iteration. Instead of training and continuously iterating multiple separate models, the unified model can be trained once on diverse data covering all scenarios, significantly reducing training costs and enabling more efficient continuous iteration and updates.
3Device complexity
If a unified model is used to process all spoken language requests, then system complexity and training costs are reduced, but the ability to accurately identify multiple intents may be compromised
Solution Approach 1:
The unified model employs a segmentation approach by predicting multiple sub-requests from a single spoken language request, where each sub-request corresponds to a specific intent with its own intent, domain, and slot information. This allows the model to accurately identify multiple intents within a single request by breaking down the complex input into distinct, structured sub-requests, thereby maintaining high intent recognition accuracy while using a single model.
Solution Approach 2:
The unified model changes its output parameters to accommodate multiple intents by generating a list of sub-requests, each with specific intent, domain, and slot parameters. This parameter-based approach allows the model to flexibly represent multiple intents, domains, and slots without requiring separate models, maintaining measurement precision while reducing system complexity.
Data Source
AI summary
A text processing method performed by an electronic device includes: obtaining a text of a spoken language request; predicting a quantity N of sub-requests of the spoken language request and N sub-request texts in one-to-one correspondence with the N sub-requests through a prediction model by using the spoken language request text, N being a positive integer; and predicting target instruction information corresponding to each of the N sub-request texts through the prediction model based on the quantity N of sub-requests and the N sub-request texts for determining a response to the spoken language request.