NLU Model Training with Domain-Intent Context for Ambiguous Utterances
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing text embedding models fail to incorporate the context of words effectively, leading to ambiguity in natural language understanding (NLU) models, particularly in scenarios requiring nuanced understanding of phrases like 'check balance', which can refer to either checking an account balance or depositing a paper check.
Innovation Solution
The approach involves structuring training data with domains and intents, using threading, annealing, and classification to enhance context understanding, allowing NLU models to accurately categorize user utterances into relevant intents.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If existing text embedding models are used to transform natural language into vectors, then the NLU model can process utterances, but the model fails to incorporate word context effectively leading to ambiguity in understanding phrases like 'check balance'
Solution Approach 1:
The training data is segmented into multiple domains (e.g., banking, general) and further divided into intents within each domain. This hierarchical segmentation allows the NLU model to first identify the domain context, then the specific intent, thereby preserving and utilizing contextual information that would otherwise be lost in flat text embedding models.
Solution Approach 2:
The patent introduces an additional dimensional structure by organizing training data into domains and intents rather than using traditional flat text embedding. This multi-dimensional organization (domain → intent → utterance) enables the model to capture contextual relationships that single-dimensional word vectors cannot represent, resolving the ambiguity in phrases like 'check balance'.
2Reliability
If the NLU model is trained with additional context through domain and intent structuring, then the accuracy and flexibility of the model improves, but the complexity of training data structuring and model deployment increases
Solution Approach 1:
The patent performs preliminary structuring of training data into domains and intents before model training. By pre-organizing the data hierarchy and generating domain-specific training utterances in advance, the complex structuring work is done upfront, allowing the actual model training and deployment to proceed more smoothly despite the inherent complexity of the multi-dimensional structure.
Solution Approach 2:
The patent introduces domain and intent as intermediary layers between the raw utterance and the final classification. These intermediaries serve as structured representations that bridge the gap between unstructured text and structured model expectations, managing complexity by breaking down the classification task into manageable hierarchical steps.
Data Source
AI summary
The present disclosure describes an electronic device. The electronic device includes a memory and a processor. The processor is configured to provide an utterance to an NLU model, which selects an intent from the list of intents based on the utterance. The processor is also configured to cause a response to the utterance to be provided to a user based on the selected intent. The NLU may be trained by obtaining a dataset comprising one or more words, generating a list of intents based at least in part on the dataset, grouping the list of intents into one or more domains, generating a list of training utterances for each intent in the list of intents, and modifying one or more parameters of the NLU model based on the list of intents, the list of training utterances, and/or the one or more domains.