DFKI-NLP at SemEval-2024 Task 2: Towards Robust LLMs Using Data Perturbations and MinMax Training

Bhuvanesh Verma, Lisa RaithelNAACL 2024

TL;DR

This paper tackles the challenge of making Large Language Models (LLMs) more robust for Natural Language Inference (NLI) in the critical biomedical domain, specifically for clinical trial reports. They achieve this by combining data perturbations (introducing variations like acronyms and numerical changes) with a MinMax training strategy that forces the model to focus on difficult examples, leading to improved performance and robustness.

What Problem Does This Paper Address?

Large Language Models (LLMs) have shown impressive capabilities across various NLP tasks, including reasoning. However, they are prone to "shortcut learning," relying on superficial patterns rather than true understanding. This is particularly problematic in high-stakes fields like biomedicine, where errors can have serious consequences. The SemEval-2024 Task 2 focused on NLI for clinical trial reports (CTRs), highlighting the need for LLMs that can reliably handle numerical reasoning, domain-specific terminology, and subtle semantic nuances within these complex documents. Existing LLMs often struggle with these challenges, making their deployment in critical applications risky.

Methods

The authors employed a two-pronged approach to enhance LLM robustness:

  • Data Perturbations: They augmented the NLI4CT dataset with two types of perturbations:
  • Acronym-based: Introduced variations by replacing acronyms with their full forms and vice-versa, and also generated negative instances by using dissimilar long forms.
  • Numerical: Modified numerical values and introduced semantic alterations to numerical comparisons, effectively creating new instances with flipped labels.
  • MinMax Training: Inspired by previous work, they integrated an auxiliary model alongside a fine-tuned LLM (Mistral-7B). This auxiliary model is designed to amplify the loss on "hard" examples (those where the primary model struggles), forcing the LLM to pay more attention to these challenging instances and thus improving its overall robustness.
  • Pre-training: The models were pre-tuned on the MedNLI dataset to familiarize them with clinical data before fine-tuning on the NLI4CT dataset.
  • Instruction Fine-tuning: They experimented with various instruction templates, ultimately framing the NLI task as a text generation problem where the model predicts "Yes" or "No" based on the premise and hypothesis.

Datasets

  • NLI4CT Dataset: The primary dataset used for training and evaluation, derived from Clinical Trial Reports (CTRs). It contains instances labeled as entailment or contradiction.
  • MedNLI Dataset: Used for pre-training the models to adapt them to the clinical domain.

Key Insights and Ideas

  • Data Perturbations are Crucial: Introducing specific perturbations, especially those mimicking real-world variations like acronyms and numerical shifts, significantly helps in training more robust models.
  • MinMax Training for Focus: The MinMax strategy effectively directs the LLM's attention to its weaknesses, leading to improved performance on challenging examples and better generalization.
  • Domain Adaptation Matters: Pre-training on a related clinical dataset (MedNLI) before fine-tuning on the target task (NLI4CT) provides a valuable boost.
  • "Hard" Examples Hold the Key: The most significant improvements often come from explicitly training the model to handle instances it previously found difficult.
  • Adverse Events are Tricky: The "Adverse Events" section of CTRs proved to be particularly challenging for models, suggesting it requires more sophisticated reasoning.

Key Results

The proposed system achieved a competitive ranking in the SemEval-2024 Task 2, securing 11th place in macro F₁ score, 12th in Faithfulness, and 19th in Consistency out of 31 participants. Specifically:

  • The MinMax trained model showed a slight improvement in Consistency and a significant improvement in Faithfulness compared to the base model, indicating better handling of semantically altering interventions.
  • While overall F₁ scores didn't always increase with perturbations, specific interventions like "Definitions" saw notable improvements from acronym-based perturbations.
  • The "Adverse Events" section yielded the highest F₁ score (0.73), while the "Eligibility" section had the lowest (0.66).
  • Among interventions, "Numerical_contradiction" achieved the highest score (0.93), while "Definition" had the lowest (0.63).

Limitations and Future Work

  • Semantic-Preserving Interventions: The system still struggled with semantic-preserving interventions, indicating room for further improvement in understanding subtle linguistic nuances.
  • Numerical Reasoning Challenges: Despite efforts, the model continued to exhibit difficulties with numerical reasoning, particularly in scenarios involving complex calculations and numerous variables.
  • Potential for Biases: The authors acknowledge the risk of sensitive content in the data and the potential for Mistral models to inherit biases from their pre-training corpora.
  • Future Directions: The authors suggest focusing future research on improving robustness by specifically targeting challenging sections of CTRs and exploring methods to mitigate issues with numerical reasoning and semantic-preserving interventions. They also highlight the need for further analysis of potential mislabeled instances and model confusion between cohorts and trials.

Summary updated

BibTeX
@inproceedings{verma2024dfki,
  author = {Bhuvanesh Verma and Lisa Raithel},
  title = {{DFKI-NLP at SemEval-2024 Task 2: Towards Robust LLMs Using Data Perturbations and MinMax Training}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2024},
}
Figure