Automated Dialog Hardening via Adversarial Test Generation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional techniques fail to effectively generate test cases for automated dialog systems that address adversarial input leading to misclassification, particularly in entity detection, and lack human-interpretable explanations for users to make informed decisions during system hardening.
Innovation Solution
An automated dialog hardening system generates test cases and human-interpretable explanations by analyzing conversation workspaces, using techniques like tf-idf representations, LIME for feature importance, and paraphrase generation to identify misclassifications, and suggests modifications to intents, entities, and dialog flows to prevent misclassification.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional testing techniques are used for automated dialog systems, then the system can operate with basic functionality, but the system remains vulnerable to misclassification from adversarial inputs
Solution Approach 1:
The system performs preliminary actions by automatically generating test cases from training data before deployment, modifying portions of training data to create adversarial test inputs that expose misclassification vulnerabilities in advance
Solution Approach 2:
The system serves itself by automatically generating test cases, identifying misclassifications, and providing explanations without requiring external testing tools or manual test case creation, thereby improving reliability while managing complexity
2Loss of information
If automated testing without explanations is implemented, then testing can be performed quickly, but users cannot understand or address misclassifications effectively
Solution Approach 1:
The system introduces an intermediary explanation layer that translates complex model decisions into human-interpretable explanations, mediating between the automated testing process and user understanding without significantly increasing analysis time
3Measurement precision
If manual analysis of misclassifications is performed, then detailed understanding can be achieved, but the process becomes time-consuming and inefficient
Solution Approach 1:
The system replaces manual mechanical analysis with automated computational processes that generate test cases and explanations, achieving both high precision in identifying misclassification causes and improved testing efficiency through automation
Data Source
AI summary
Methods, systems and computer program products for automated testing of dialog systems are provided herein. A computer-implemented method includes receiving information pertaining to a given conversation workspace of an automated dialog system and identifying test case inputs to the automated dialog system, the test case inputs comprising user input for the given conversation workspace that has portions thereof modified and which the automated dialog system maps to a different intent and/or a different entity relative to the user input. The method further includes generating human-interpretable explanations of mappings of portions of the test case inputs to the different intent and/or entity, generating suggestions for modifying intents, entities and dialog flows of the given conversation workspace such that the test case inputs map to the same intent and/or the same entity as their corresponding user input, and outputting the suggestions and the human-interpretable explanations to a user.