Language Model Building via Categorized Sentence Mining
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for building language models face challenges in accurately covering multiple vertical fields due to the difficulty in incorporating uncommon and unpopular words, leading to reduced accuracy in speech recognition systems.
Innovation Solution
A method involving categorized sentence mining, text classification, and training of template-based, class-based, and lexicon-based language models to improve field-specific language models without increasing the number of data samples, using techniques like weighted finite state transduction and confidence score calculation for integrated speech recognition results.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If data mining is performed on a dictionary to obtain initial language training samples covering many fields, then the language model can cover multiple vertical fields, but the difficulty of access to the dictionary increases and the difficulty of building language models increases
Solution Approach 1:
The patent segments the language model building process into multiple stages: initial sample acquisition, interpolation with fixed weights, and subsequent refinement using field-specific training samples. This segmentation allows the system to handle multiple fields systematically without requiring access to comprehensive dictionaries for all fields simultaneously, thereby reducing the complexity of building language models while maintaining multi-field coverage.
2Adaptability or versatility
If interpolation method with fixed weights is used to expand initial speech training samples to cover multiple vertical fields, then the language model becomes a balanced multi-vertical-field language model, but the interpolated training samples fail to include terminology, uncommon words and unpopular words in a vertical field
Solution Approach 1:
The patent applies preliminary action by first performing interpolation with fixed weights to create a balanced multi-field language model, then subsequently refining it using field-specific training samples that contain terminology, uncommon words, and unpopular words. This two-stage approach ensures both balanced coverage and high precision for vertical field identification, resolving the contradiction between balance and accuracy.
3Adaptability or versatility
If the capacity of the dictionary is increased to cover more fields, then the initial language training samples can cover more fields, but the difficulty of access to the dictionary increases
Solution Approach 1:
The patent extracts and utilizes field-specific training samples from various sources for each vertical field, rather than relying on a single comprehensive dictionary. This extraction approach allows the system to cover multiple fields by gathering specialized samples for each field independently, avoiding the need to access and manage a large-capacity dictionary that would be difficult to obtain and maintain.
Data Source
AI summary
A method includes: acquiring data samples; performing categorized sentence mining in the acquired data samples to obtain categorized training samples for multiple categories; building a text classifier based on the categorized training samples; classifying the data samples using the text classifier to obtain a class vocabulary and a corpus for each category; mining the corpus for each category according to the class vocabulary for the category to obtain a respective set of high-frequency language templates; training on the templates for each category to obtain a template-based language model for the category; training on the corpus for each category to obtain a class-based language model for the category; training on the class vocabulary for each category to obtain a lexicon-based language model for the category; building a speech decoder according to an acoustic model, the class-based language model and the lexicon-based language model for any given field, and the data samples.