Intention Identification Model Learning via Negative Data and Ensemble
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing human-machine dialog systems face accuracy issues in intention identification due to insufficient or inaccurate training data and improper base models, leading to misidentification and poor user experience.
Innovation Solution
The introduction of negative data generation and multi-task learning using similar voice skills to enrich base models, combined with ensemble learning, to improve the accuracy of intention identification models.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If only positive training data is used for model learning, then the learning process is simple, but misidentification occurs and accuracy decreases
Solution Approach 1:
The patent applies preliminary anti-action by generating negative training data that pre-counters potential misidentification errors. The system proactively creates counter-examples (negative data) that represent cases where misidentification might occur, allowing the model to learn from these counter-examples before actual deployment. This prevents misidentification by preparing the model in advance with contrasting examples that highlight the boundaries between different intentions.
2Quantity of substance
If training data is insufficient or inaccurate, then data collection cost is low, but model accuracy is severely affected
Solution Approach 1:
The patent applies copying by generating synthetic negative training data through text generation models. Instead of requiring extensive manual collection of real negative examples, the system creates copies or simulations of negative cases by transforming positive examples into negative ones through controlled text generation. This allows the model to learn from a much larger and more diverse set of training examples without the high cost of manual data collection.
3Productivity
If a single base model is used for learning, then the learning process is fast, but model accuracy is limited
Solution Approach 1:
The patent applies merging by combining multiple base models into an ensemble learning system. The system trains and integrates predictions from multiple different base models (e.g., different text classification algorithms, different neural network architectures) to make a final intention identification decision. This ensemble approach leverages the strengths of different models while compensating for their individual weaknesses, achieving higher accuracy than any single model could achieve alone.
4Measurement precision
If negative data is generated and similar skill training data is introduced, then model accuracy improves, but learning complexity increases
Solution Approach 1:
The patent applies universality by creating a multi-functional learning framework that handles multiple tasks simultaneously. The same ensemble learning system and negative data generation mechanism are applied across different skills and domains, making the complex system reusable and manageable. The framework can process positive and negative data, train multiple base models, and integrate them using a unified approach that works across various intention identification tasks, reducing the perceived complexity through standardized procedures.
Data Source
AI summary
An intention identification model learning method includes receiving positive data that corresponds to a first skill, generating, based on the positive data that corresponds to the first skill, negative data that corresponds to the first skill, determining a second skill similar to the first skill, obtaining data that corresponds to each second skill, generating a second base model based on the data that corresponds to the second skill and a first base model stored on the server, and performing learning based on the second base model, the positive data, and the negative data that correspond to the first skill, and generating an intention identification model.