Semantic Class Generator for ASR Language Model Expansion
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current automatic speech recognition (ASR) systems face limitations in generating semantically similar sentences, which are essential for improving language models and user query interfaces, as they rely on restricted vocabularies and lack the ability to produce grammatically correct variations of input utterances.
Innovation Solution
A system comprising a semantic class generator to identify semantically similar words and a sentence verifier to produce grammatically correct sentences, using ontologies or dictionaries, which can be integrated into existing statistical language models for ASR and natural language processing applications.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If current ASR systems use restricted vocabularies and existing sentence databases, then system complexity is reduced, but the ability to generate semantically similar sentences is limited
Solution Approach 1:
The system segments the sentence generation task into distinct functional modules: a semantic class generator that identifies semantically similar words using ontologies, a sentence generator that assembles candidate sentences, and a sentence verifier that checks grammatical correctness. This modular segmentation enables versatile sentence generation while managing system complexity through organized functional decomposition.
Solution Approach 2:
The patent introduces intermediary components between the input utterance and the final sentence output: the semantic class generator acts as an intermediary that expands vocabulary using ontology relationships, and the sentence verifier serves as an intermediary filter that ensures grammatical correctness. These intermediaries enable the system to generate diverse sentences without directly managing the full complexity of language generation.
2Measurement precision
If the system generates diverse sentence variations, then language model accuracy improves, but the computational resources and processing time increase
Solution Approach 1:
The system performs preliminary actions by pre-building ontology structures and semantic class relationships before sentence generation is needed. The semantic class generator maintains pre-computed semantic relationships that can be quickly queried during sentence generation, reducing real-time processing requirements while still enabling diverse sentence variation for improved language model accuracy.
Solution Approach 2:
The sentence verifier applies grammatical checking selectively to specific candidate sentences rather than all possible variations. By focusing verification resources on the most promising candidates generated by the sentence generator, the system achieves adequate grammatical correctness for improved language model accuracy without verifying every possible sentence variation, thus reducing overall processing time.
3Adaptability or versatility
If the system expands vocabulary beyond existing limitations, then sentence diversity increases, but the reliability of grammatical correctness may decrease
Solution Approach 1:
The sentence verifier serves as a critical intermediary between the diverse sentence candidates generated from expanded vocabulary and the final output. It acts as a quality gate that checks grammatical correctness, ensuring that even sentences generated from newly expanded vocabulary terms meet acceptable grammatical standards before being added to the language model, thus maintaining reliability while enabling sentence diversity.
Solution Approach 2:
The system implements feedback through the sentence verifier that evaluates grammatical correctness of generated sentences and provides information back to the sentence generator. This feedback mechanism allows the system to learn from grammatical errors in diverse sentence variations and adjust future generation to maintain higher grammatical reliability while preserving vocabulary expansion benefits.
Data Source
AI summary
A system and method are described for generating semantically similar sentences for a statistical language model. A semantic class generator determines for each word in an input utterance a set of corresponding semantically similar words. A sentence generator computes a set of candidate sentences each containing at most one member from each set of semantically similar words. A sentence verifier grammatically tests each candidate sentence to determine a set of grammatically correct sentences semantically similar to the input utterance. Also note that the generated semantically similar sentences are not restricted to be selected from an existing sentence database.