The way you talk to a chatbot may be more important than you think
Oscar Wong/Getty Images
Speaking to an AI chatbot in a less formal language, as many people do, reduces the accuracy of its responses. This means that either we need to be more linguistically rigorous when using a chatbot, or that the AI needs to be trained to better adapt to informality.
Amazon's Fulei Zhang and Zhou Yu looked at how people initiate conversations with human agents compared to a large language model (LLM) chatbot assistant. They used the Claude 3.5 Sonnet model to evaluate conversations on a number of factors and found that people interacting with chatbots used less precise grammar and were less polite than when speaking to humans. They also used a slightly narrower vocabulary.
For example, Claude's scores found that human interactions were 14.5 percent more polite and formal than interactions with chatbots, 5.3 percent more fluid, and 1.4 percent more lexically diverse.
“Users adapt their linguistic style in human-to-LLM communications, producing messages that are shorter, more direct, less formal, and grammatically simpler,” write the authors, who did not respond to an interview request, in an article about their work. “This behavior is likely shaped by the mental models of LLM chatbot users.[s] as less socially sensitive or less capable of subtle interpretation.”
But it turns out that this informality also has a downside. In the second evaluation, the researchers trained an AI model called Mistral 7B on 13,000 real conversations between people and used it to interpret 1,357 real messages sent to AI chatbots. They annotated each conversation in both datasets with an “intent” taken from a limited list, summarizing what the user was trying to do in each case. But since Mistral's AI was trained on human-to-human conversations, the pair found that the AI struggled to correctly indicate intent in conversations with chatbots.
Zhang and Yu then tried different strategies to improve the AI's understanding of Mistral. First, they used Claude's AI to rewrite users' more succinct messages into humane prose, and used them to fine-tune Mistral's model. This reduced the accuracy of intent labels by 1.9 percent compared to default responses.
They then used Claude for “minimal” rewrites that were shorter and blunter (for example, “Paris next month. Hotel flights?” to ask about travel and accommodation options for an upcoming trip), but this reduced Mistral's accuracy by 2.6 percent. An alternative, “rich” rework with more formal and varied language also resulted in a 1.8% decrease in accuracy. And just by training the Mistral model on both minimal and extended rewrites, performance improved by 2.9%.
Noah Giansiracusa from Bentley University in Massachusetts says he's not surprised that people talk to bots differently than they do to humans, but it shouldn't necessarily be avoided.
“The fact that people interact with chatbots differently than they do with other humans is temptingly touted as a disadvantage of a chatbot—but I would argue that this is not the case, which is good as long as people know they are talking to bots and adapt their behavior accordingly,” says Giansiracusa. “I think it’s more useful than obsessively trying to bridge the gap between human and bot.”
Topics:
