Researchers Find New Potential in Coronavirus Chatbots
Artificial intelligence researchers are developing data to train coronavirus chatbots. The scientists tested multiple AI models on discussions about the coronavirus between doctors and patients with the goal of creating “meaningful medical dialogue” about COVID-19with the chatbot.
The research grew out of the increasing need for telemedicine services buttressed by AI. Chatbots built by different medical facilities and AI developers are becoming a common method for people to assess their own health and whether they should make an appointment with a doctor to check for infection by the coronavirus. Scientists at UC San Diego, Carnegie Mellon University, and UC Davis tested different AI chatbots by training them on the coronavirus conversations, called CovidDialog, in both English and Chinese. They will publish the results in an upcoming paper. The data came from a mix of online healthcare forums, 603 English consultations, and 1,088 Chinese consultations. The patient described their medical condition, and then the doctor and patient would discuss potential diagnoses and treatment.
The researchers applied the conversations to Google’s Transformer architecture, OpenAI’s GPT language model, and the BERT-GPT encoder-decoder architecture. To make sure the AI wasn’t too constrained by the coronavirus data to be useful, all of the models used transfer learnings, where they were pre-trained on much larger datasets before the CovidDialog data was brought in. The results for all of the models were potentially promising, but while the Transformer and GPT models struggled with coherence, the BERT-GPT model performed quite well. The demand for AI that can help take some of the load off of medical professionals is great enough to make further research worthwhile, according to the researchers.
“To address the large imbalance between the surging need of consultations from citizens and the severe shortage of medical professionals available to provide online consultation services, it is highly valuable to develop intelligent dialogue systems which act as virtual doctors to provide COVID-related consultations to people. These virtual doctors can greatly ease the burden of human doctors and timely address the concerns of the public,” the authors wrote. “Experiments demonstrate that the models trained on CovidDialog datasets are promising in generating clinically meaningful consultations about COVID-19.”
While academic research continues, COVID-19 chatbots are mushrooming worldwide. That includes both text and voice form. For instance, India and the UK have each published a WhatsApp chatbot to answer questions about the pandemic. In the U.S., state governments have also started asking voice app developers like Voicify to design ways to communicate with citizens about COVID-19 through Alexa and Google Assistant. For healthcare providers, companies like Orbita are building interactive voice and text chatbots. Medical professionals and policymakers are all looking for ways to slow the rush of calls to hospitals and doctors, whether that means integrating Hyro’s free coronavirus-focused version of its virtual assistant or adapting Microsoft’s template for the same purpose. The biggest tech companies are following suit on their own platforms, with Google Assistant offering pandemic tips, or Apple’s Siri and Amazon’s Alexa voice assistants providing their own COVID-19 questionnaire to assess potential infection. The study makes a good case for why these are only the beginning of the medical chatbot renaissance.