Two-Stage Intent Recognition for Fast, Accurate Query Understanding
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing intent recognition methods in artificial intelligence, particularly using large language models, face challenges with complex user queries due to high resource consumption, slow response times, and inaccurate intent identification, especially when queries are rich and colloquial or express multiple intents in a single sentence.
Innovation Solution
A two-stage approach using a large model for information enhancement followed by a deep learning model with fewer parameters for intent recognition, where the large model preprocesses the query to enhance information and the deep learning model quickly identifies candidate intents, reducing resource usage and improving accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a large language model is used for intent recognition, then accuracy is improved, but response time increases and resource consumption increases
Solution Approach 1:
The patent divides the intent recognition system into two segments: a large language model for information enhancement and a lightweight deep learning model for actual intent classification. This segmentation allows the large model to process only the necessary preprocessing tasks while the lightweight model handles the time-critical classification, thereby improving response time while maintaining accuracy.
Solution Approach 2:
The large language model performs information enhancement preprocessing before the intent recognition stage. By enriching and structuring the input data in advance, the large model enables the lightweight model to make accurate decisions faster, effectively preparing the data beforehand to reduce processing time during critical inference.
2Measurement precision
If a large language model is used for intent recognition, then accuracy is improved, but power consumption increases
Solution Approach 1:
The system segments computational tasks between a large language model and a lightweight deep learning model. The large model performs energy-intensive information enhancement offline or on server, while the lightweight model handles energy-constrained edge device inference, thereby reducing overall power consumption while maintaining accuracy.
Solution Approach 2:
The patent creates a simplified copy of the intent recognition capability using a lightweight deep learning model that replicates the essential classification function without the computational overhead of the large language model, enabling energy-efficient deployment on resource-constrained devices.
3Speed
If traditional deep learning models are used for intent recognition, then response time is reduced, but accuracy decreases for complex queries
Solution Approach 1:
The large language model performs preliminary information enhancement that transforms complex, colloquial queries into structured, semantically enriched representations. This preprocessing enables the fast lightweight model to accurately interpret complex queries that would otherwise be difficult to classify, thereby improving accuracy without sacrificing speed.
Solution Approach 2:
The system changes the parameter representation of input queries by using the large language model to generate enhanced semantic representations. This transformation of input parameters enables the lightweight model to achieve higher accuracy on complex queries while maintaining its computational efficiency and fast response time.
4Measurement precision
If more labeled data is collected for fine-tuning, then model accuracy is improved, but data collection costs increase
Solution Approach 1:
The patent uses the large language model to generate synthetic enhanced training data that copies and transforms existing labeled data into augmented variations. This approach improves model accuracy by providing more diverse training examples without the need to collect and annotate additional real-world data, thereby reducing data collection costs.
Data Source
AI summary
An intent recognition method includes obtaining question information, inputting the question information into a first model for information enhancement processing to obtain a first processing result, inputting the first processing result into a second model for intent recognition processing to obtain a second processing result that includes at least one candidate intent corresponding to the question information, and determining, based on the second processing result, a target intent corresponding to the question information. The first model is a large model and the second model is a deep learning model with an intent recognition speed faster than the large model for same question information.