Unbalanced Training Data Batching for Chatbot Bias Reduction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Chatbot systems trained with unbalanced data exhibit biased behaviors, leading to inefficient or incorrect interactions with users, which can waste resources and cause customer complications.
Innovation Solution
The system utilizes an intent distribution to generate a distributed and unbiased training batch, ensuring that each output category is utilized equally by the intent classifier, thereby reducing biases inherent in previous random-sampling methods.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If random sampling is used to generate training batches, then the training process is simple and fast, but the trained model exhibits biased behavior towards overrepresented categories
Solution Approach 1:
The patent applies local quality by making the sampling probability dependent on the specific category characteristics. Instead of uniform random sampling, the system adjusts the sampling probability for each category based on its representation in the training data, thereby creating non-uniform local properties in the sampling process that counteract the global bias towards overrepresented categories
Solution Approach 2:
The patent changes the sampling parameter from uniform probability to category-dependent probability. By modifying the sampling distribution parameter to account for category frequencies, the system transforms the sampling process to achieve balanced category representation in training batches while maintaining computational efficiency
2Measurement precision
If training data is provided with high representation of certain contexts, then the chatbot learns those contexts well, but it exhibits biased behavior and performs poorly on underrepresented contexts
Solution Approach 1:
The patent applies dynamics by making the training batch composition adaptive rather than static. The sampling probability dynamically adjusts based on category representation, ensuring that underrepresented categories receive adequate attention in each training batch while maintaining overall training efficiency and model performance across all contexts
3Use of energy by moving object
If the chatbot is trained with unbalanced data, then training resources are consumed efficiently, but biased interactions occur wasting resources and causing customer complications
Solution Approach 1:
The patent converts the harmful effect of data imbalance into a benefit by using the imbalance information to guide the sampling process. Instead of treating unbalanced data as a problem to be ignored, the system leverages the frequency information to adjust sampling probabilities, thereby transforming the potential harm of bias into a useful signal for achieving balanced training
Data Source
AI summary
The present disclosure relates to chatbot systems, and more particularly, to batching techniques for handling unbalanced training data when training a model such that bias is removed from the trained machine learning model when performing inference. In an embodiment, a plurality of raw utterances is obtained. A bias eliminating distribution is determined and a subset of the plurality of raw utterances is batched according to the bias-reducing distribution. The resulting unbiased training data may be input into a prediction model for training the prediction model. The trained prediction model may be obtained and utilized to predict unbiased results from new inputs received by the trained prediction model.