Text Formatter Using ML Sub-word Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing technologies struggle to properly format raw text extracted from audio conversations, leading to potential misinterpretation and legal issues, as seen in cases where a missing comma can alter the meaning of a contract.
Innovation Solution
A method and system that utilize machine-learning models to convert raw text into formatted text by generating sub-words, applying formatting operations such as punctuation, capitalization, and entity recognition, and combining these operations to produce text that accurately reflects the intended meaning of the speaker.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional text formatting methods are used on raw text from audio conversations, then the formatting process is simple and fast, but the formatting accuracy and meaning preservation deteriorate
Solution Approach 1:
The text formatting process is divided into multiple specialized models: a punctuation model that adds appropriate punctuation marks, a capitalization model that capitalizes proper nouns and sentence beginnings, and an entity recognition model that identifies and formats entities. Each model focuses on a specific formatting aspect, improving overall accuracy while keeping individual model complexity manageable.
Solution Approach 2:
Multiple specialized formatting models are combined into an integrated text formatting system. The punctuation model, capitalization model, and entity recognition model work together in sequence to produce properly formatted text, achieving high formatting accuracy through the synergistic combination of multiple specialized components.
2Manufacturing precision
If machine-learning models are used to format text, then the formatting quality and meaning preservation improve, but the processing time and computational resources increase
Solution Approach 1:
The text formatting task is segmented into multiple independent models (punctuation, capitalization, entity recognition) that can process text in parallel or sequence. This segmentation allows each model to be optimized for its specific function, improving overall formatting quality while enabling more efficient processing compared to a single monolithic model.
Solution Approach 2:
The system performs preliminary text processing by first transcribing audio to raw text, then immediately applying multiple formatting models in sequence. The punctuation model adds punctuation first, followed by capitalization and entity recognition, ensuring that formatting operations are performed in the optimal order to maintain meaning while improving efficiency.
3Productivity
If simple formatting rules are applied to raw text, then the processing speed is fast, but the ability to preserve speaker intent and context deteriorates
Solution Approach 1:
The machine-learning models are trained on large datasets of properly formatted text, allowing them to learn from examples of correct punctuation, capitalization, and entity formatting. During processing, the models use contextual information from surrounding words to make informed formatting decisions, preserving speaker intent while maintaining fast processing speeds through efficient neural network inference.
Solution Approach 2:
The system uses multiple formatting models with different parameters and functions: the punctuation model adjusts punctuation parameters, the capitalization model adjusts case parameters, and the entity recognition model adjusts entity boundary parameters. These parameter changes allow the system to preserve context and intent while processing text efficiently through optimized model inference.
Data Source
AI summary
Methods, systems, and computer programs are presented for formatting raw text. One method includes an operation for accessing raw text comprising words corresponding to one or more sentences. The raw text is lowercase text without any punctuation. Further, the method includes operations for creating a plurality of sub-words corresponding to the raw text, and for generating, by a machine-learning (ML) model, an output for each sub-word based on the created sub-words. The output for each sub-word indicates a formatting operation for the corresponding sub-word. The method further includes an operation for generating, based on the formatting operations in the outputs for the sub-words, formatted text corresponding to the raw text. The formatted text is text with correct grammar, proper punctuation, and proper capitalization according to a meaning of words spoken by a speaker associated with the raw text.