Dialogue Intention Recognition Model Few-Shot Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for recognizing dialogue intentions face challenges in accuracy due to overfitting caused by the small scale of labeled data, especially in cold-start scenarios, leading to low recognition results and domain shift issues in few-shot learning.
Innovation Solution
A method utilizing a trained intention recognition model that generates sample sentence vectors and category vectors based on input sentences and labeled supporting sentences, reducing dependence on the scale of supporting sentences through few-shot learning and incorporating feature extraction and attention mechanisms to improve accuracy and efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If labeled data is used as the basis for dialogue intention recognition, then the recognition can be performed, but the small scale of labeled data causes overfitting and reduces recognition accuracy
Solution Approach 1:
The patent introduces a pre-trained language model as an intermediary to bridge the gap between limited labeled data and accurate intention recognition. The model leverages extensive pre-training on unlabeled corpus to learn general language patterns, then adapts to specific intention recognition tasks with minimal labeled data through few-shot learning, effectively using the pre-trained knowledge as a mediator to overcome data scarcity
Solution Approach 2:
The patent applies preliminary action by pre-training the language model on large-scale unlabeled corpora before performing intention recognition. This pre-training phase establishes foundational language understanding and contextual representation capabilities, which are then transferred to the specific task with limited labeled data, eliminating the need to start from scratch with small datasets
2Adaptability or versatility
If few-shot learning is used to reduce dependence on supporting sentence scale, then the model can work with limited data, but domain shift issues arise leading to low recognition results
Solution Approach 1:
The patent employs parameter changes by dynamically adjusting the number of shots (k) in few-shot learning based on domain characteristics and task requirements. The system can adaptively select different amounts of supporting sentences for different domains or intention types, optimizing the balance between adaptability to new domains and maintaining recognition accuracy under domain shift conditions
3Productivity
If traditional intention recognition methods are used, then the process can be completed, but overfitting occurs due to small labeled data scale
Solution Approach 1:
The patent applies copying by utilizing the knowledge and patterns learned from large-scale pre-training corpora and copying them to the target intention recognition task. The pre-trained model captures general language understanding that can be transferred and applied to specific intention recognition scenarios, allowing the system to benefit from extensive training data without requiring proportional labeled data for each specific task
Data Source
AI summary
A method for recognizing a dialogue intention, an electronic device and a storage medium are provided. The solution includes: obtaining a sample sentence with an intention category to be predicted and a plurality of supporting sentences each labeled with an intention category; and inputting the sample sentence and the plurality of supporting sentences into a trained intention recognition model, in which the trained intention recognition model is configured to generate a sample sentence vector corresponding to the sample sentence and category vectors corresponding to the plurality of supporting sentences based on the sample sentence and the plurality of supporting sentences, calculate matching degrees between the sample sentence vector and the category vectors, and obtain a predicted intention category of the sample sentence based on the matching degrees.