TL;DR
This paper introduces a novel framework to automatically discover new rules for generating Natural Language Inference (NLI) training data. By combining Out-of-Distribution (OOD) detection with BERT-based clustering, they identify gaps in existing NLI rules and discover four new ones, leading to improved NLI model performance and a more scalable approach to data augmentation.
What Problem Does This Paper Address?
Natural Language Inference (NLI) is a crucial task in NLP, requiring models to understand logical relationships between sentences. However, training effective NLI models for new domains is hindered by the high cost and effort of creating domain-specific labeled data. While previous work attempted to automate this using predefined sentence transformation rules, the existing set of 15 rules proved insufficient to capture the full diversity of natural language patterns. This paper addresses this limitation by proposing a method to automatically discover new transformation rules, thereby expanding the coverage and quality of automatically generated NLI training data.
Methods
The core of the proposed method lies in identifying NLI premise-hypothesis pairs that are "out-of-distribution" with respect to the existing 15 transformation rules. This is achieved through a combination of OOD detection techniques:
- Maximum Softmax Probability (MSP): A baseline method that uses the highest softmax probability as an indicator of confidence.
- Temperature Scaling (TS): Calibrates softmax probabilities to better align with actual model accuracy, mitigating overconfidence.
- Input Preprocessing (IP): Modifies input embeddings to enhance the separation between in-distribution and out-of-distribution data.
These OOD pairs are then clustered using BERT embeddings and k-means clustering. By analyzing these clusters, particularly those with high cohesion and consistent explanations, the authors derive new transformation rules. To further automate this process, they also explore using Large Language Models (LLMs) with Chain-of-Thought (CoT) prompting to generate and validate new rule candidates. The discovered rules are then used with LLMs and CoT prompting to generate high-quality NLI training data.
Datasets
- SNLI (Stanford Natural Language Inference) dataset: Used for training and evaluating NLI models, as well as for analyzing rule distributions.
- e-SNLI dataset: Utilized for its human-annotated natural language explanations, which were found to be beneficial for clustering.
Key Insights and Ideas
- OOD Detection as a Gap Identifier: The paper cleverly uses OOD detection not just for model robustness, but as a mechanism to find areas where existing NLI rules are insufficient. This is a novel application of OOD techniques.
- Clustering for Rule Discovery: Grouping OOD premise-hypothesis pairs reveals underlying patterns that can be formalized into new transformation rules. The insight that explanation-based clustering yields better results is significant.
- LLMs for Scalable Rule Discovery: Leveraging LLMs with CoT prompting offers a promising avenue for automating the laborious manual analysis required for rule discovery, making the process more scalable and efficient.
- Distribution-Aware Augmentation Matters: The finding that augmenting data to reflect the natural distribution of rules in the SNLI dataset leads to better performance, especially on smaller datasets, highlights the importance of realistic data generation.
Key Results
- Discovery of Four New Rules: The authors successfully identified four new transformation rules: Role Generalization (RG), Contextual Augmentation (CA), Visual Specification (VS), and Emotion Inference (EI).
- Performance Improvements: Augmenting the SNLI dataset with data generated by the new rules led to consistent performance improvements across various dataset sizes, with gains of +0.85%p on 2k samples and +0.15%p on 550k samples.
- Distribution-aware Augmentation Effectiveness: This strategy showed superior performance across all dataset sizes, particularly for small datasets (2k).
- Automated Rule Discovery Validation: The automated pipeline using LLMs successfully rediscovered three of the four manually identified rules, demonstrating its reliability and efficiency.
- CoT Explanations as a Scalable Alternative: Using LLM-generated explanations (CoT-Ex) for clustering proved to be a viable and scalable alternative to human-annotated explanations.
Limitations and Future Work
- Design Choices Dependency: The automated rule discovery process still relies on several design choices, such as similarity thresholds and filtering criteria, which could be further explored.
- Quality and Diversity of LLM-Generated Data: A comprehensive validation of the quality and diversity of LLM-generated data compared to human-annotated data is still needed, and potential biases introduced by LLMs remain underexplored.
- Generalizability of CoT Explanations: The consistency and reliability of CoT explanations across different domains, datasets, and languages require further rigorous validation.
- Dataset Specificity: The current augmentation strategy has only been evaluated on the SNLI dataset, and further studies are needed to establish its generalizability to other NLP tasks and datasets.
Summary updated