ESR6: Alignment in human-machine spoken interaction

PhD Fellow: Greta Gandolfi

I am a first-year PhD student at the University of Edinburgh, under the supervision of Professor Martin Pickering.

My research starts from one simple assumption: that the nature of language is social. Therefore, I try to study the relationship between language, brain and environment going beyond the classic monadic approach often employed by the cognitive sciences.

Specifically, I am studying what can explain our tendency, as human beings, to adapt to our interlocutors’ language use. I am particularly curious about the role that social factors, background and prior knowledge can play in conversations.

Before being involved in the COBRA project, I studied philosophy at the San Raffaele University in Milan and cognitive sciences at CIMeC (Center for Mind / Brain Sciences) in Trento, with particular focus on language and multimodal interaction. The non-linearity of my study path allowed me to explore what I am passionate about, that is how knowledge spreads between individuals and groups of people, through language.

First, I analysed the relationship between language and knowledge from the point of view of social epistemology. Then, I continued through the lens of computational sociolinguistics. I am now conducting experiments in the framework of the psychology of language.

I truly believe in the power of interdisciplinarity, in the values ​​of open science, in knowledge exchange and sharing, in the importance of dynamic environments. That’s why I’m excited to be a part of this project.

Objectives:

In the context of a dialogue game in which interlocutors describe objects to each other as part of a task, ESR6 will investigate the extent to which people align with each other and with a natural-language generation system, and how this impacts on task success. Starting from a simple referential communication game, in which objects may be referred to by means of different terms, ESR6 will move on to more naturalistic tasks such as determining a route through a complex environment. By manipulating the performance of the generation system (e.g., use of preferred vs dispreferred terms), we can determine what characteristics enhance alignment and task success. Manipulations will be directly relevant to alignment (e.g., whether the system always aligns with the person) as well as related to other aspects of the task (e.g., when and how often backchannels are used). ESR6 will also manipulate the human-likeness of the system.


Expected results:

Alignment is expected to lead to task success (a deceptively simple claim that needs carefully-controlled experimentation to test). Specifically, we expect a direct relationship between alignment of the generation system and alignment of participants (i.e., if the system aligns with you, you will tend to align with it). The effects of other manipulations are likely to be less direct; for example, backchannels may lead to more succinct descriptions and it may be easier to align on these descriptions.

Based in Edinburgh, UK

Full-time three-year contract, starting September 2020

PhD enrolment at: University of Edinburgh

Main supervisor’s institution: University of Edinburgh

Main supervisor: Prof Martin Pickering

Secondments:

  • DAVI, Puteaux: training on using real-time dialogue systems in experimental settings (5 months);
  • Furhat Robotics, Stockholm: to test the system with a physical agent/robot (5,5 months).

Co-supervisors’ institutions:

  • DAVI, Puteaux, France
  • Furhat Robotics, Stockholm, Sweden


Scroll to top