Adversarial Distraction for NLI Syntax Robustness
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Natural language inference classifiers trained on existing datasets make errors due to their inability to understand how syntax affects meaning, leading to inadequate performance in identifying entailment relationships.
Innovation Solution
The method involves generating adversarial distractions by creating pairs of examples with differing syntactic structures but matching information, which are used to retrain the neural network models, enhancing their robustness to syntax changes through techniques like person reversal and passive voice transformation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If natural language inference classifiers are trained on existing datasets, then they achieve near human level performance on standard tasks, but they make errors due to inability to understand how syntax affects meaning
Solution Approach 1:
The patent applies preliminary action by proactively generating adversarial distraction examples with reversed person references before training the classifier. These synthetic examples are created in advance to teach the model about syntax-meaning relationships, preventing syntax-related errors before they occur in actual inference tasks.
Solution Approach 2:
The patent implements preliminary anti-action by introducing adversarial examples that deliberately contain syntax manipulations (person reversals) opposite to correct patterns. These counter-examples are designed to counteract the model's tendency to ignore syntax, thereby preventing future syntax-related classification errors.
2Reliability
If adversarial distraction training is applied to teach syntax, then robustness to syntax changes is enhanced, but training data processing complexity increases
Solution Approach 1:
The patent applies copying by creating synthetic adversarial examples that replicate the structure of existing training data but with modified person references. Instead of manually creating complex training scenarios, the system copies existing sentences and applies systematic transformations (person reversals) to generate the augmented training set.
Solution Approach 2:
The patent implements parameter changes by systematically modifying specific parameters of existing sentences (person reference indices) to generate adversarial examples. This approach transforms a small number of original sentences into multiple training variants by changing only the person reference parameters, rather than creating entirely new complex examples.
Data Source
AI summary
A computer-implemented method and system are provided for teaching syntax for training a neural network based natural language inference model. The method includes selectively performing, by the hardware processor, person reversal on a set of hypothesis sentences, based on person reversal prevention criteria, to obtain a first training data set. The method further includes enhancing, by the hardware processor, a robustness of the neural network based natural language inference model to syntax changes by training the neural network based natural language inference model on original training data combined with the first data set.