AI-driven assistants are gaining significant traction in the modern world. AI assistants, like Alexa and Siri, need to exhibit apt social behavior to engage users effectively. Inspired by this, researchers at Uber and Carnegie Mellon developed a machine learning model to introduce social language into the assistant’s responses.
The research was centered on the customer service domain, typically a single-use case. In this case, a few drivers signed up with a ride-sharing provider like Uber. The researchers first studied the relationship between the customer service representative’s use of friendly language to drivers’ responses and the completion of their first trip. Further, they developed a machine learning model for an assistant to incorporate a language generation and social language understanding component.
Based on the data set obtained after training the model on nearly 233,000 messages, the responses had parameters indicating how generally polite they were as analyzed by human evaluators. Furthermore, the researchers used automated and human-driven techniques to evaluate their model’s messages’ politeness and positivity. They discovered an interesting fact, while the model could preserve the intended meaning of the message and vary the politeness level of the responses, it was less successful in maintaining overall positivity. The researchers feel this may be because of the difference between what they intended to measure and evaluate and what they actually measured and analyzed.
The researchers believe that this model can help the customer service representatives deliver a quick response and help them stick to best practices while responding. This model is another step forward in the direction where many organizations prefer chatbots compared to the human voice.