Automatic Agent Feedback for Unhuman Response Breakdowns
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Human-machine interaction models, such as automatic agents, often respond in an unhuman fashion when encountering unfamiliar dialog, leading to breakdowns and inability to solve user problems effectively.
Innovation Solution
The system improves automatic agents by providing feedback on mistakes through an improvement mode, where a sympathetic user corrects the agent's behavior by selecting alternative replies from a list of options, updating the machine learning policy based on these corrections, and retraining the agent using the updated dialog sequences.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Extent of automation
If the automatic agent uses a machine learning policy to generate replies, then it can respond automatically to user inputs, but it may respond in an unhuman fashion when encountering unfamiliar dialog
Solution Approach 1:
The system implements feedback by detecting breakdowns in the dialog flow and using them to adjust the machine learning policy. When the agent encounters unfamiliar dialog and generates unhuman responses, the system detects this breakdown and uses the feedback to retrain the policy, allowing the agent to learn from its mistakes and improve future responses.
Solution Approach 2:
The system performs preliminary action by training the machine learning policy on a corpus of dialog samples before deployment. This pre-training enables the agent to have a baseline level of performance on common dialog patterns, and subsequent feedback from breakdowns further refines this preliminary knowledge.
2Adaptability or versatility
If the automatic agent responds to unfamiliar dialog, then it can handle new situations, but it may produce unhuman responses leading to breakdown
Solution Approach 1:
The system uses feedback from breakdown detection to continuously improve the agent's adaptability. When the agent handles unfamiliar dialog and produces unhuman responses, the breakdown detection mechanism captures this feedback and uses it to adjust the machine learning policy, enabling the agent to learn from unfamiliar situations and improve its adaptability over time.
Solution Approach 2:
The system changes parameters of the machine learning policy based on feedback from breakdowns. By adjusting the policy parameters through retraining on updated dialog sequences, the system enables the agent to adapt its behavior when encountering unfamiliar dialog patterns, improving both adaptability and response quality.
3Reliability
If the system provides feedback to improve the automatic agent, then the agent can correct its behavior, but it requires additional processing and training time
Solution Approach 1:
The system discards problematic dialog sequences that lead to breakdowns and recovers by training on corrected versions. Instead of continuously training on all data, the system identifies and focuses on the specific breakdown cases, discarding the faulty patterns and recovering through targeted retraining on corrected dialog sequences, thereby reducing overall training time while improving reliability.
4Adaptability or versatility
If the automatic agent uses a machine learning policy, then it can learn from interactions, but it may breakdown when encountering dialog it has not seen before
Solution Approach 1:
The system implements feedback by detecting when the agent's learned responses break down in unfamiliar situations. This feedback loop allows the agent to stabilize by learning from its failures, combining adaptability from machine learning with stability through continuous correction of breakdown cases.
Solution Approach 2:
The system provides beforehand cushioning by training the machine learning policy on a comprehensive corpus of dialog samples before deployment. This pre-training cushions the agent against encountering completely unfamiliar dialog, and subsequent feedback from any remaining breakdowns provides additional cushioning to further stabilize performance.
Data Source
AI summary
An automatic agent may be improved through feedback. A user input may be received through a user interface. A plurality of current utterance variables may be obtained by tokenizing the user input. The automatic agent may execute a machine learning policy to generate a reply to the user input based on the plurality of current utterance variables. A different reply may be obtained in response to an indication that the reply will lead to a breakdown, wherein the breakdown comprises an unhuman response from the automatic agent according to the machine learning policy. The machine learning policy may be adjusted based on the plurality of current utterance variables and the different reply.