Brief Enquiries

How Can You Sensitise Siri, Ms Gašić?

By Jan Berndorff

Thanks to modern dialogue systems like Alexa and Siri, computers have become our everyday helpers: on voice command they deliver information, operate stereo equipment, switch the lights on and even tell jokes. But they do it without a trace of emotion – a normal conversation is impossible. Milica Gašić from Serbia wants to change that.

Milica Gašić
Milica Gašić (Photo: Humboldt Foundation / Mats Karlsson)

“Siri, who is Helene Fischer for heaven’s sake?” The response from the digital assistant in our smartphone is quite prosaic: a German pop singer. She ignores the underlying irritation in our voice. Irony, sadness and enthusiasm are also lost on her. Nor is she a dab hand at deep conversation.

Because that requires empathy – the conversation partner must be able to sense emotions and respond to them. “A system could recognise emotions by the sound of the voice or choice of words,” says Milica Gašić.

To deduce emotions from the choice of words – this is what Gašić wants to teach the systems: apart from encyclopaedic knowledge, the AI researcher feeds in recordings of conversations between real people. A statistical model evaluates the connections between choice of words and the interlocutor’s response. This enables the system to develop answers that are emotionally appropriate. “Conversations should become more human; the systems shouldn’t only provide information but also impart a positive feeling,” says Gašić.

published in Humboldt Kosmos 110/2019

The Sofja Kovalevskaja Award Winner Professor Milica Gašić spent three months working at Saarland University and now holds the Chair in Dialog Systems and Machine Learning at Heinrich Heine University Düsseldorf.