Validating training data of classifiers

Abstract

A method, a computer program product, and a computer system for evaluation and modification of training data for classifiers. A computer system applies an each piece of training data of an each classto multiple classifiers. The computer system performs evaluation and validation of the training data for the each class and defines at least one representative class for the each piece of the trainingdata of the each class if classification precision of the each piece of the training data is greater than a predetermined threshold. The computer system modifies the training data, based on results of the evaluation and the validation of the training data for the each class. The computer system performs evaluation and validation of the training data for representative classes. The computer systemmodifies the training data, based on results of the evaluation and the validation of the training data for the representative classes.

Description

technical field [0001] The present invention relates generally to validating training data for classifiers, and more particularly to evaluating and modifying training data for classifiers of automated response systems. Background technique [0002] Large volumes of data including semi-structured and unstructured data are known as big data. In recent years, technologies that apply big data to business activities and medical services have attracted attention. In particular, technologies such as cognitive computing and deep learning have been applied in these fields. Cognitive computing provides the ability to recognize natural language, speech and image data. By learning from large amounts of data, deep learning has greatly improved the accuracy of machine learning. [0003] An automatic answering system is a system applying those technologies. Auto-response systems provide answers to questions in natural language, such as speech or typed text. In such systems, by using c...

AI Technical Summary

Problems solved by technology

However, in the case of the above-mentioned automatic answering system, the data part is the problem statement written in natural language, so there are the following problems: (1) It is not easy to collect the problem statement; (2) The intention of the problem statement collected as training data is usually Has content dedicated to a specific domain, so it is necessary to individually verify and determine intent by experts in each domain, etc.

Images