Voice Intent Modeling for Predicate-Free Utterance Processing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing speech recognition systems struggle when a user's utterance lacks a verb (predicate), leading to incorrect intent determination and subsequent inappropriate content delivery.
Innovation Solution
An electronic device detects failures in intent determination due to missing predicates and generates alternative intent models based on the input text, improving content delivery accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If speech recognition systems use traditional intent determination models, then processing is simple and fast, but intent determination fails when utterances lack verbs (predicates)
Solution Approach 1:
The system performs preliminary analysis of the utterance structure before intent determination. When a verb is detected as missing, the system proactively generates alternative intent models based on the noun phrases present in the utterance, rather than attempting to process the incomplete utterance with the original model. This preliminary detection and model generation approach ensures reliable intent determination even for incomplete utterances.
Solution Approach 2:
The system changes the parameter of intent determination by dynamically selecting different models based on utterance characteristics. When an utterance lacks a verb, the system switches from using the standard intent determination model to using generated intent models that are specifically trained for noun-only utterances. This parameter change (model selection) allows the system to maintain high accuracy without permanently increasing complexity.
2Measurement precision
If the system generates alternative intent models for every failed intent determination, then intent interpretation accuracy improves, but processing time increases
Solution Approach 1:
The system performs preliminary classification of utterances to identify those that are likely to fail intent determination (e.g., utterances without verbs). By detecting these problematic utterances beforehand and generating appropriate models only for them, the system avoids unnecessary processing time for successful utterances while ensuring accurate model generation for failed cases.
Solution Approach 2:
The system applies partial action by generating alternative intent models only for the subset of utterances that fail initial intent determination, rather than generating models for all utterances. This selective approach ensures high accuracy for problematic cases while minimizing the time cost of model generation for straightforward utterances.
3Ease of manufacture
If the system provides content services based on traditional speech recognition, then service delivery is straightforward, but wrong content is provided when utterances are incomplete
Solution Approach 1:
The system implements feedback by using the result of intent determination to guide content service selection. When intent determination fails or produces uncertain results, the system uses this feedback to trigger alternative model generation, which then produces more accurate intent interpretations that lead to correct content delivery. This feedback loop ensures reliability without complicating the overall content delivery process.
Solution Approach 2:
The generated intent models act as intermediaries between the incomplete utterance and the content service. Instead of directly providing wrong content based on faulty intent recognition, the intermediary models translate the incomplete utterance into accurate intent representations, which then correctly select appropriate content services.
Data Source
AI summary
An electronic device includes a microphone, a memory and at least one processor. The at least one processor is configured to acquire utterance data corresponding to a voice of a user through the microphone, determine an intent and provide content to the user based on the intent. When intent determination fails, then one or more models are updated using text obtained from the utterance data. The intent determination may fail when there is no verb (predicate) in the user utterance. The models are updated by searching for named entities and determining domains to be used for the model updates. The domains are determined based on categories. The categories are found using a named entity search (NES). Examples of categories are music artists, music albums, movie titles, TV program channels, video clip channels, radio programs, and podcast titles.