Backpropagation Training for Dialog System Error Propagation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional dialog systems trained with clean data struggle to handle errors introduced within their pipeline, leading to diminished user experience due to propagation of inaccuracies across machine learning models like ASR, semantic parser, and TTS subsystems.
Innovation Solution
The use of backpropagation to train machine learning models in a dialog system by converting seed data through multiple formats, comparing the results to the original data, and using the differences as training signals to improve model accuracy and reduce error propagation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional dialog systems are trained with clean data, then the training process is simple and efficient, but the systems struggle to handle errors introduced within their pipeline and errors propagate across models
Solution Approach 1:
The patent implements a feedback mechanism where the system converts data through multiple models (ASR, semantic parser, TTS) and uses the output to generate training signals that feed back into the training process. This allows the system to learn from its own errors and improve reliability without requiring completely separate training datasets.
Solution Approach 2:
The dialog system trains itself by using its own components to convert seed data and generate training signals. The system serves its own training needs by internally converting data through ASR, semantic parsing, and TTS subsystems, eliminating the need for external manual annotation processes.
2Adaptability or versatility
If multiple machine learning models (ASR, semantic parser, TTS) are integrated into a pipeline, then the system can process speech through multiple formats, but errors are introduced and propagated across the pipeline
Solution Approach 1:
The system uses the output from subsequent models in the pipeline (semantic parser and TTS) to generate training signals for earlier models (ASR). This feedback loop allows each model to learn from errors that propagate through the pipeline, maintaining accuracy despite the multi-format processing capability.
Solution Approach 2:
The system performs preliminary conversions of seed data through the entire pipeline before using the results for training. By pre-converting the data and comparing outputs at different stages, the system identifies and corrects errors before they affect final system performance.
3Quantity of substance
If the system converts seed data through multiple formats using ASR, semantic parser, and TTS subsystems, then comprehensive training data is generated, but computational resources and time are consumed
Solution Approach 1:
The existing ASR, semantic parser, and TTS subsystems are made multi-functional by using them both for their primary operations and for generating training data. This universal usage allows the system to expand its training data quantity without adding separate dedicated data generation infrastructure.
Solution Approach 2:
The system generates its own training data by internally converting seed data through the pipeline. This self-service approach eliminates the need for external data generation processes and reduces overall training time by leveraging existing computational resources already deployed in the system.
Data Source
AI summary
Techniques described herein use backpropagation to train one or more machine learning (ML) models of a dialog system. For instance, a method includes accessing seed data that includes training tuples, where each training tuple comprising a respective logical form. The method includes converting the logical form of a training tuple to a converted logical form, by applying to the logical form a text-to-speech (TTS) subsystem, an automatic speech recognition (ASR) subsystem, and a semantic parser of a dialog system. The method includes determining a training signal by using an objective function to compare the converted logical form to the logical form. The method further includes training the TTS subsystem, the ASR subsystem, and the semantic parser via backpropagation based on the training signal. As a result of the training by backpropagation, the machine learning models are tuned work effectively together within a pipeline of the dialog system.