Georgia Tech researchers say non-English speakers shouldn’t rely on chatbots like ChatGPT to provide valuable healthcare advice.
A team of researchers from the College of Computing at Georgia Tech has developed a framework for assessing the capabilities of large language models (LLMs).
Ph.D. students Mohit Chandra and Yiqiao (Ahren) Jin are the co-lead authors of the paper Better to Ask in English: Cross-Lingual Evaluation of Large Language Models for Healthcare Queries.
Their paper’s findings reveal a gap between LLMs and their ability to answer health-related questions. Chandra and Jin point out the limitations of LLMs for users and developers but also highlight their potential.
Their XLingEval framework cautions non-English speakers from using chatbots as alternatives to doctors for advice. However, models can improve by deepening the data pool with multilingual source material such as their proposed XLingHealth benchmark.
“For users, our research supports what ChatGPT’s website already states: chatbots make a lot of mistakes, so we should not rely on them for critical decision-making or for information that requires high accuracy,” Jin said.
“Since we observed this language disparity in their performance, LLM developers should focus on improving accuracy, correctness, consistency, and reliability in other languages,” Jin said.
Using XLingEval, the researchers found chatbots are less accurate in Spanish, Chinese, and Hindi compared to English. By focusing on correctness, consistency, and verifiability, they discovered:
- Correctness decreased by 18% when the same questions were asked in Spanish, Chinese, and Hindi.
- Answers in non-English were 29% less consistent than their English counterparts.
- Non-English responses were 13% overall less verifiable.
XLingHealth contains question-answer pairs that chatbots can reference, which the group hopes will spark improvement within LLMs.
The HealthQA dataset uses specialized healthcare articles from the popular healthcare website Patient. It includes 1,134 health-related question-answer pairs as excerpts from original articles.
LiveQA is a second dataset containing 246 question-answer pairs constructed from frequently asked questions (FAQs) platforms associated with the U.S. National Institutes of Health (NIH).
For drug-related questions, the group built a MedicationQA component. This dataset contains 690 questions extracted from anonymous consumer queries submitted to MedlinePlus. The answers are sourced from medical references, such as MedlinePlus and DailyMed.
In their tests, the researchers asked over 2,000 medical-related questions to ChatGPT-3.5 and MedAlpaca. MedAlpaca is a healthcare question-answer chatbot trained in medical literature. Yet, more than 67% of its responses to non-English questions were irrelevant or contradictory.
“We see far worse performance in the case of MedAlpaca than ChatGPT,” Chandra said.
“The majority of the data for MedAlpaca is in English, so it struggled to answer queries in non-English languages. GPT also struggled, but it performed much better than MedAlpaca because it had some sort of training data in other languages.”
Ph.D. student Gaurav Verma and postdoctoral researcher Yibo Hu co-authored the paper.
Jin and Verma study under Srijan Kumar, an assistant professor in the School of Computational Science and Engineering, and Hu is a postdoc in Kumar’s lab. Chandra is advised by Munmun De Choudhury, an associate professor in the School of Interactive Computing.
The team will present their paper at The Web Conference, occurring May 13-17 in Singapore. The annual conference focuses on the future direction of the internet. The group’s presentation is a complimentary match, considering the conference's location.
English and Chinese are the most common languages in Singapore. The group tested Spanish, Chinese, and Hindi because they are the world’s most spoken languages after English. Personal curiosity and background played a part in inspiring the study.
“ChatGPT was very popular when it launched in 2022, especially for us computer science students who are always exploring new technology,” said Jin. “Non-native English speakers, like Mohit and I, noticed early on that chatbots underperformed in our native languages.”
School of Interactive Computing communications officer Nathan Deen and School of Computational Science and Engineering communications officer Bryant Wine contributed to this report.
