Query Rewriting for Natural Language Understanding Errors
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing speech processing systems face issues with errors in automatic speech recognition (ASR) and natural language understanding (NLU) that lead to user frustration and friction, such as misinterpretation of user inputs, resulting in undesired responses and delays.
Innovation Solution
A system that pre-trains a query embedder using historical dialog session data and fine-tunes it with rephrase pairs to generate alternative representations of user inputs, reducing friction by ensuring desired responses.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of operation
If automatic speech recognition and natural language understanding are used to enable voice-based user control, then human-computer interaction is improved, but errors in interpretation occur leading to undesired responses and user frustration
Solution Approach 1:
The system implements a feedback mechanism where the generated alternative representation is evaluated against the original user input using a trained model. The model predicts whether the alternative representation will result in a desired response, and this feedback loop allows the system to select or generate more accurate alternative representations, thereby improving interpretation reliability while maintaining ease of operation
Solution Approach 2:
The system performs preliminary actions by pre-processing the original user input to generate an alternative representation before the main NLU processing occurs. This pre-processing step, involving query embedding and alternative generation using historical dialog data, prepares a corrected or improved version of the input that reduces the likelihood of interpretation errors in subsequent processing stages
2Adaptability or versatility
If speech recognition and natural language understanding processing techniques are combined, then speech-based user control is enabled, but misinterpretation of user inputs occurs resulting in user frustration
Solution Approach 1:
The system introduces an intermediary component - the query embedder and alternative representation generator - that sits between the original user input and the NLU processing. This intermediary uses historical dialog session data to generate alternative representations that are less likely to be misinterpreted, thereby reducing user frustration while preserving the adaptability of speech-based control
Solution Approach 2:
The system applies beforehand cushioning by using historical dialog session data to pre-train the query embedder model. This preparation in advance creates a buffer against potential misinterpretations by equipping the system with learned patterns from historical data, which helps cushion against errors and reduce user frustration before they occur
Data Source
AI summary
A system is provided for reducing friction during user interactions with a natural language processing system, such as voice assistant systems. The system determines a pre-trained model using dialog session data corresponding to multiple user profiles. The system determines a fine-tuned model using the pre-trained model and a fine-tuning dataset that corresponds to a particular task, such as query rewriting. The system uses the fine-tuned model to process a user input and determine an alternative representation of the input that can result in a desired response from the natural language processing system.