TL;DR
This paper presents a three-module system for Natural Language Inference on Clinical Trial Reports (NLI4CT), a challenging task due to the complexity and potential for LLM hallucinations. Their novel approach leverages LLM-generated intermediate labels to fine-tune each module, achieving a respectable rank in the SemEval-2024 Task 2 competition and offering interesting insights into the nuances of clinical text inference.
What Problem Does This Paper Address?
The paper tackles the critical challenge of applying Large Language Models (LLMs) to real-world problems, specifically in the domain of clinical trials. While LLMs have revolutionized Natural Language Inference (NLI), their tendency to "hallucinate" (generate factually incorrect information) poses a significant obstacle, especially in fields like biomedical applications where accuracy is paramount. The NLI4CT dataset, introduced for SemEval-2024 Task 2, specifically addresses this by requiring entailment judgments between clinical trial reports and associated statements, a task that often necessitates a series of inferential steps rather than direct information retrieval. This complexity makes it a prime candidate for exploring robust NLI systems that can mitigate LLM hallucinations.
Methods
The authors developed a three-module system:
- Query Generation (QG) Module: This module generates specific questions based on a given statement and evidence to verify the statement's truthfulness.
- Query Answering (QA) Module: This module takes the generated queries and the evidence to produce answers.
- Aggregation (AG) Module: This module synthesizes the evidence, queries, and answers to make the final entailment prediction.
A key innovation is the use of LLM-generated intermediate labels. Instead of directly training on the NLI4CT dataset, they first used LLMs (Gemini-Pro and Mistral-7B-Instruct-v0.2) with few-shot and zero-shot prompting to generate query-answer pairs for each statement. These generated pairs were then used to fine-tune the individual modules. This approach aims to guide the LLMs towards more accurate and grounded reasoning by providing them with structured, intermediate steps. They experimented with various combinations of pre-trained and fine-tuned modules, including custom prompts tailored to specific sections of the clinical trial reports.
Datasets
The primary dataset used is the NLI4CT dataset, specifically designed for SemEval-2024 Task 2. This dataset comprises clinical trial reports, statements derived from these reports, and their corresponding entailment labels. The dataset contains 999 clinical trial reports, with a training set of 1700 statements and a development set of 200 statements. The test set for the competition included 5500 statements. The authors also generated their own intermediate dataset of query-answer pairs using LLMs for fine-tuning purposes.
Key Insights and Ideas
- LLM-generated intermediate labels are crucial: The authors found that fine-tuning modules with LLM-generated query-answer pairs significantly improved performance compared to zero-shot or few-shot prompting alone. This highlights the value of structured intermediate reasoning steps for LLMs.
- Semantic-preserving interventions are harder to detect: The experiments suggest that interventions that subtly alter meaning without drastically changing it are more challenging for the system to identify than those that make more obvious semantic shifts.
- Negative to positive entailment is trickier: The system found it more difficult to correctly infer entailment when a statement was negatively entailed (i.e., contradicted the evidence) and the system predicted positive entailment, compared to the reverse.
- Aggregation module sensitivity: Switching the type of AG module (e.g., from pre-trained to custom pre-trained) had a more consistent impact on performance metrics like faithfulness and F1 score than changes in the QA module.
Key Results
The system submitted for the SemEval-2024 Task 2 competition achieved a rank of 19th-24th across different metrics. Their best experimental configuration (experiment 8) scored:
- F1-score: 0.659
- Consistency: 0.579
- Faithfulness: 0.595
While not top-tier, the authors highlight that their best F1 score was achieved in experiment 3 (0.666), and their best consistency and faithfulness were in experiment 2 (0.657 and 0.781, respectively). The error analysis revealed that the system often struggles with inferring numerical results, suggesting a need for more specialized handling of such data.
Limitations and Future Work
The authors acknowledge several limitations:
- Resource constraints: Their fine-tuning experiments were limited to relatively small open-sourced LLMs due to resource limitations.
- Data size: The smaller size of their LLM-generated intermediate dataset compared to the original NLI4CT training set might have impacted performance.
- Context window: The limited context size of the LLMs used for prompting QA and AG modules restricted the number of examples that could be included.
Future work suggestions include:
- Integrating rule-based systems or code generation: To better handle numerical reasoning, which is a current weakness.
- Filtering evidence: More effectively filtering evidence input based on cohorts and trials relevant to the query.
- Exploring larger LLMs: Extending the approach to larger and more capable LLMs.
- Increasing example diversity: Using LLMs to generate a wider variety of examples with larger context sizes.
Summary updated