Proxy Tags for Intent Recognition Model Portability
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current language modeling approaches face challenges in training effective natural language processing systems due to the need for exhaustive listing of all possible surface forms, leading to non-portable training samples and statistical biases, especially in scenarios with limited natural language examples.
Innovation Solution
The use of proxy tags and surface form variants within training examples allows for the creation of modular and swappable pieces, enabling the building of semantic analysis models without requiring explicit listing of all possible phrases, and facilitates faster training and adaptation across different conversational domains.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If all possible surface form variants are explicitly listed in training examples, then the language model can recognize equivalent variations, but the training data becomes non-portable and requires exhaustive listing of every alternative phrasing
Solution Approach 1:
The patent introduces proxy tags as intermediary elements that represent sets of surface form variants. Instead of directly listing all possible phrasing variations in training examples, the system uses proxy tags (e.g., @person_name, @location) as mediators that map multiple surface forms to a single semantic representation. This resolves the contradiction by maintaining recognition accuracy through the proxy tag system while eliminating the need for exhaustive training data listing.
Solution Approach 2:
The patent creates training examples with universal structures that can represent multiple specific cases. By using proxy tags that can stand for any instance within a category (e.g., @person_name can represent any person's name), a single training example becomes universally applicable to all variants within that category. This eliminates the need for separate training examples for each surface form variant while maintaining the ability to recognize all equivalent variations.
2Reliability
If a fixed corpus of text is used for training, then the language model can be built with traditional methods, but increasing corpus size to improve performance becomes challenging due to the need for explicit alternative phrasing
Solution Approach 1:
The patent changes the parameter representation from surface form (specific words and phrases) to semantic form (proxy tags). By transforming the training data parameterization, the system achieves better model performance without requiring an increase in the quantity of training examples. A single training example with proxy tags effectively represents multiple surface form variants, making the training process more efficient and scalable.
3Reliability
If every alternative phrasing is provided in training examples, then the classifier can be reliably trained, but the training process becomes overly fitted to lexical features and loses portability
Solution Approach 1:
The patent uses proxy tags as reusable templates that can be copied and applied across different domains and contexts. Instead of copying and adapting specific surface form variants for each new application, the system copies the proxy tag structure and populates it with domain-specific instances. This maintains reliable classifier training through consistent semantic representations while enabling easy portability across different conversational domains by simply changing the proxy tag instantiations.
Data Source
AI summary
A computer-implemented method for building a semantic analysis model. In one embodiment, the computer-implemented method includes creating proxy tags comprising a set of surface form variants. The computer-implemented method creates training examples comprising a combination of terminal tokens and at least one of the proxy tags. The computer-implemented method builds the semantic analysis model using the training examples.