Huatuo-26M, a Large-scale Chinese Medical QA Dataset (2305.01526v1)

Abstract: In this paper, we release a largest ever medical Question Answering (QA) dataset with 26 million QA pairs. We benchmark many existing approaches in our dataset in terms of both retrieval and generation. Experimental results show that the existing models perform far lower than expected and the released dataset is still challenging in the pre-trained LLM era. Moreover, we also experimentally show the benefit of the proposed dataset in many aspects: (i) trained models for other QA datasets in a zero-shot fashion; and (ii) as external knowledge for retrieval-augmented generation (RAG); and (iii) improving existing pre-trained LLMs by using the QA pairs as a pre-training corpus in continued training manner. We believe that this dataset will not only contribute to medical research but also facilitate both the patients and clinical doctors. See \url{https://github.com/FreedomIntelligence/Huatuo-26M}.

Overview of "Huatuo-26M, a Large-scale Chinese Medical QA Dataset"

The research paper introduces Huatuo-26M, a comprehensive Chinese medical question-answering (QA) dataset that includes 26 million QA pairs. This dataset is distinguished by its unparalleled scale, aiming to meet the data demands of pre-trained LLMs (PLMs) in medical NLP applications. The authors emphasize the difficulties of applying PLMs to medical fields, highlighting the urgent need for abundant and high-quality data. This dataset is expected to enhance the capabilities of PLMs in the medical domain.

Dataset Composition and Acquisition

The Huatuo-26M dataset comprises QA pairs derived from diverse sources: online medical consultation records, medical encyclopedias, and existing medical knowledge bases. The data was meticulously collected, cleaned, and de-duplicated, resulting in a dataset that not only surpasses existing medical QA datasets in scale but also offers a wide variety of questions and professional answers.

• Online Medical Consultation Records: The largest portion of the dataset was obtained from online platforms where patients seek advice from medical professionals. This contributed to a significant number of colloquial questions paired with expert responses.
• Medical Encyclopedias and Knowledge Bases: Additional data was harvested from medical encyclopedias and structured knowledge bases, complementing the data with more systematic and formal medical information.

Benchmarking and Evaluation

Huatuo-26M was assessed using classical retrieval and generative models, marking a rigorous benchmark in the domain:

• Retrieval Models: Sparse (BM25, DeepCT) and dense (DPR) retrieval models were evaluated, demonstrating that existing models face challenges when applied to this dataset, primarily due to the intricate medical knowledge required.
• Generative Models: The authors fine-tuned T5 and GPT2 models on Huatuo-26M data. Despite significant improvements post fine-tuning, the generated answers still indicate room for improvement, reflecting the complexity and diversity of the dataset content.

Practical and Theoretical Implications

The dataset has several potential applications:

1. Transfer Learning: Models pretrained on Huatuo-26M can generalize well to other medical QA datasets, demonstrating its efficacy as a pre-training resource for transfer learning.
2. Retrieval-Augmented Generation (RAG): Leveraging the dataset as an external knowledge base significantly improves text generation quality in RAG setups, highlighting the dataset's value in integrating PLMs with non-parametric memory.
3. Continued Pre-training: Using Huatuo-26M as a pre-training corpus enhances existing models' performance in medical NLP tasks, as evidenced by improvements on the CBLUE medical benchmark.

Conclusions and Future Directions

Huatuo-26M stands as a landmark contribution to Chinese medical NLP, providing a substantial resource that aligns with the pressing need to integrate AI into healthcare. The dataset offers extensive opportunities for future research, notably in enhancing LLM-based QA systems' ability to reason with domain-specific knowledge. However, the authors caution that the dataset may contain inaccuracies and that generation technologies' current limitations pose risks in real-world medical applications. Future improvements could involve expanding the dataset to other languages and systematically curating the data to ensure higher accuracy and reliability.

Overall, Huatuo-26M presents both opportunities and challenges, charting a path forward for increased AI involvement in healthcare through more effective and accessible LLMs.