Dependency Tree Augmentation for Sentence Well-Formedness
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing natural language processing methods for judging sentence well-formedness primarily use binary rating scales, which are inadequate for multiple fluency or grammaticality rating scales, and face challenges due to data scarcity, especially for low-resource languages, requiring innovative data augmentation techniques.
Innovation Solution
A dependency tree-based data augmentation method that applies a dependency parser to generate a tree for a sentence, removes nodes based on a predetermined removal ratio and rating score, and generates a partial tree to create rated sentences for training, utilizing a universal dependency tree structure and token importance to simulate varying degrees of sentence ill-formality.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If binary classification is used for sentence well-formedness judgement, then the classification process is simple, but it cannot support multiple fluency or grammaticality rating scales
Solution Approach 1:
The patent segments the binary classification task into multiple rating categories by introducing an uncertainty threshold mechanism. The model outputs probability distributions over multiple rating classes, and a segmentation threshold determines whether to output a specific rating or indicate uncertainty, thus supporting multiple rating scales while maintaining computational efficiency.
Solution Approach 2:
The patent introduces dynamic threshold adjustment for uncertainty classification. The threshold for determining uncertain sentences is not fixed but can be adjusted based on model confidence and application requirements, allowing the system to adapt between binary and multi-class classification modes dynamically.
2Measurement precision
If human annotation is used to create training data, then the quality of training data is high, but the process is time-consuming and costly
Solution Approach 1:
The patent implements self-service data augmentation by enabling the model to generate its own training data through uncertainty sampling. Sentences that the model finds uncertain are automatically selected and used for targeted augmentation, reducing reliance on expensive human annotation while maintaining data quality through model-guided selection.
Solution Approach 2:
The patent changes the parameter of data quality from being solely dependent on human annotation to being enhanced by model-generated uncertain samples. By identifying and augmenting uncertain predictions, the system improves training data quality through a different mechanism that does not require human time investment.
3Reliability
If rule-based noise injection is used for data augmentation, then expert knowledge can be incorporated, but it requires expert knowledge for each language and is heuristics-based
Solution Approach 1:
The patent replaces the mechanical expert-knowledge-based rule system with a data-driven neural network model. Instead of manually crafting rules requiring expert linguistic knowledge, the system learns patterns automatically from data, substituting human expert mechanisms with automated machine learning approaches.
4Adaptability or versatility
If statistics noise learned injection is used for data augmentation, then learning from corpuses is enabled, but it is not available for many languages especially low-resource ones
Solution Approach 1:
The patent performs preliminary data augmentation by generating synthetic uncertain samples before model training. This preliminary action creates sufficient training data for low-resource languages without requiring large annotated corpora, enabling the model to learn from artificially constructed but linguistically plausible examples.
Solution Approach 2:
The patent uses copying by generating synthetic training examples that replicate the structure and characteristics of scarce annotated data. Through uncertainty-based sampling and augmentation, the system creates multiple copies and variations of limited training data, effectively expanding the available training material for low-resource languages.
Data Source
AI summary
A computer-implemented method, a computer program product, and a computer system for dependency tree-based data augmentation for sentence well-formedness judgement. A computer applies a dependency parser to generate a dependency tree for a sentence. The computer removes one or more nodes in the dependency tree, according to a removal ratio for a predetermined rating score. The computer generates, from the dependency tree, a partial tree for the sentence. The computer outputs a rated sentence based on the partial tree. The rated sentence is used as training data.