Iterative Data Cleaning for Chatbot Model Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional chatbot data cleaning methods are often specific to particular domains or tasks, cannot be generalized, and rely heavily on manual work, leading to increased labor costs and poor accuracy due to the use of uncleaned conversation data, which degrades the training of chatbot models.
Innovation Solution
A method involving iterative data cleaning where an initial sample set of question-answer pairs is processed to determine successive subsets based on match probability thresholds, with each model iteration improving the sample quality by removing less reliable data, ultimately enhancing chatbot model accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Extent of automation
If manual data cleaning methods are used, then labor costs increase and accuracy decreases, but automated methods lack generality and are domain-specific
Solution Approach 1:
The patent creates a universal data cleaning framework that can handle multiple domains and tasks through iterative model training. The system uses a base model that is repeatedly trained on filtered subsets of data, allowing it to adapt to different domains while maintaining a consistent automated cleaning process. This resolves the contradiction by making the automated method domain-general rather than domain-specific.
Solution Approach 2:
The patent implements dynamic adaptability through iterative model retraining. The cleaning model is continuously updated based on the filtered data subsets, allowing the system to adapt to different domains and tasks dynamically. This dynamic approach enables the automated method to be versatile across different applications while maintaining high accuracy.
2Reliability
If uncleaned conversation data is used for training, then model training degrades, but cleaning processes are too labor-intensive
Solution Approach 1:
The patent implements self-service data cleaning where the system automatically filters and cleans training data using the trained model without human intervention. The model identifies and removes low-quality samples autonomously through iterative refinement, eliminating the need for labor-intensive manual cleaning while ensuring high training quality.
Solution Approach 2:
The patent uses feedback loops where the model performance is continuously evaluated and used to refine the data cleaning process. The system feeds back the model's predictions and performance metrics to adjust the filtering criteria in subsequent iterations, automatically improving both cleaning quality and model training reliability without manual intervention.
3Measurement precision
If iterative model training with subset filtering is implemented, then data quality improves, but computational complexity increases
Solution Approach 1:
The patent divides the training data into multiple subsets through iterative filtering, processing smaller portions of data at each stage rather than handling the entire dataset at once. This segmentation approach improves sample quality by focusing on high-confidence samples while reducing the computational burden at each iteration step.
Solution Approach 2:
The patent applies partial action by processing only the most relevant data subsets at each iteration rather than the complete dataset. By focusing computational resources on filtering and training with selected subsets rather than exhaustive processing, the system achieves high sample quality with manageable computational complexity.
Data Source
AI summary
A method includes: determining a first sample subset in an initial sample set by an initial model based on the initial sample set, wherein the initial sample set comprises a plurality of question-answer pairs, each of the plurality of question-answer pairs comprising a question and an answer; generating a first model by training the initial model with the first sample subset; determining a second sample subset in the first sample subset by the first model based on the first sample subset; generating a second model by training the first model with the second sample subset; determining, in response to at least one of the second sample subset and the second model satisfying a corresponding predetermined condition, a third sample subset of the initial sample set by the second model based on the initial sample set; and generating a third model by training the second model with the third sample subset.