Code-Mixing Training Data Generation Through Sentence Alignment
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current speech-to-text models face challenges in handling code-mixing scenarios due to the scarcity of code-mixing training data and the inability to effectively handle accents and regional dialects, leading to poor performance when encountering unfamiliar speech patterns.
Innovation Solution
A training data generating device and method that aligns and translates whole sentences across languages, performs part-of-speech tagging, and selectively merges semantic units to generate a large amount of code-mixing data, ensuring context relevance and accuracy through various screening operations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If single language training data is used, then the amount of training data is large and complete, but the neural network model cannot cope with code-mixing situations
Solution Approach 1:
The patent performs preliminary actions by translating whole sentences from one language to another before alignment. The translation step is executed in advance to create parallel language versions, which are then aligned segment-by-segment to generate code-mixing training data. This preliminary translation action enables the subsequent creation of bilingual training examples.
Solution Approach 2:
The patent segments sentences into smaller text segments (e.g., phrases or clauses) rather than treating entire sentences as single units. This segmentation allows for flexible alignment and code-mixing at the segment level, enabling the model to learn language switching patterns while maintaining grammatical coherence in the generated training data.
2Reliability
If training data is screened for specific ethnic groups, then the suitability for specific accents improves, but the amount of data is drastically reduced
Solution Approach 1:
The patent creates synthetic training data by copying and translating existing single-language training data into multiple language versions. Instead of relying on scarce real-world code-mixing recordings from specific ethnic groups, the system generates artificial code-mixing examples by aligning and mixing segments from translated versions of available data, thereby expanding data quantity while maintaining relevance to specific linguistic contexts.
3Reliability
If field-specific training data is used, then the model performs well in that specific field, but performance degrades in other fields
Solution Approach 1:
The patent creates a universal code-mixing training data generation framework that can adapt to multiple fields and domains. By segmenting sentences and aligning translations at the segment level, the system generates training data that captures general code-mixing patterns applicable across different contexts (medical, business, education, etc.), enabling the model to handle code-mixing universally rather than being specialized to a single field.
Data Source
AI summary
A training data generating device and a training data generating method are provided. The device stores first single language code data, the first single language code data corresponding to a first language. The device generates a second single language code data corresponding to each of the first single language code data based on a second language and a whole sentence translation algorithm. The second single language code data corresponding to the second language. The device aligns text segments corresponding to the first single language code data and the second single language code data. The device generates code-mixing data based on at least one valid segment position corresponding to the text segments of each of the first single language code data.