The purpose of this project is to gain a better understanding of human-human dialogue through formal/computational modelling (Piwek, 2011; 2006). It will focus one of the following three phenomena in dialogue:

1. multimodal dialogue acts (involving gesture and speech) (Piwek 2009; Piwek, Beun & Cremers, 2008),
2. generation of questions (Piwek & Boyer, 2012; Kuyten et al., 2012; Rus et al. 2010; Wyse & Piwek, 2009), or
3. adversarial dialogue moves (and emotion) (Pluss, 2014; Pluss, Piwek & Power, 2011).

The main aim of the project will be to construct a computational model of one of these phenomena. This will involve building algorithms for automatically generating dialogue acts that exhibit the relevant phenomenon (for example, a model for automatically generating questions or, for instance, for generating multimodal dialogue acts).

The research may (or may not) involve developing a pair of agents that have the ability to converse with each other, with the resulting behaviour mirroring, in relevant respects, the human behaviours as we find them in naturally-occurring dialogues.

Skills and Background required

A good undergraduate degree (2.1 or above) in Computing (or equivalent), with experience in artificial intelligence, computational linguistics or natural language processing. A background in linguistics would also be suitable, if accompanied by a suitable computer skills.

References

In a number of books (e.g. Brandom, 1994; Brandom, 2000), the philosopher Robert Brandom has uncovered deep connections between dialogue games and argumentation on the one hand, and meaning on the other. Some progress has been made on making these ideas computational (see e.g. Piwek, 2014). The aim of this project is to further develop and implement these computational interpretations of Brandom's work. This may include the use of visualisations in order to get a better understanding of the models. It could also lead to development of applications (e.g., in argumentation education).

Skills and Background required

A good undergraduate degree (2.1 or above) in Computing (or equivalent), with experience in logical semantics (ideally, also proof-theoretic/inferentialist approaches) and/or dialogue modelling.

References

In the early 2020s, Large Language Models (LLMs), such as Bert, found their way into many Natural Language Processing applications. LLMs also captured the attention of the media, especially OpenAI’s GPT 2 and 3 series. And yet, the perceived real worldwide breakthrough came only in November 2022, with the release of ChatGPT. Soon traditional tests of machine intelligence, such as the Turing Test, were called into question (Biever, 2023).

By framing interactions with LLMs, that is language generators, as dialogue, it seemed that genuine conversational partners or speakers had been created. This was however not the first time that framing interaction as dialogue convinced people of the intelligence of a machine. In the 1960s, ELIZA, a simple rule-based chatbot, was able to convince some users that it was a real psychotherapist (Weizenbaum, 1966). Later research has suggested that people have a natural tendency to treat new media as real people (Reeves & Nass, 1996). This raises the question whether the framing of LLMs as conversational partners is a conceit - the chatbot conceit (Piwek, 2024) - or a genuine step towards intelligent interaction.

The project will systematically examine to what extent framing interactions with LLMs as dialogue affects people’s perceptions of the LLMs. The aim is for project’s findings to shed light on the impact of the chatbot conceit and the broader implications for the field of AI and human-computer interaction.

Skills required

A good undergraduate degree (2.1 or above) in Computing (or equivalent), with experience in artificial intelligence, computational linguistics, natural language processing or human-computer interaction.

Background Reading

Biever, Celeste (2023). ChatGPT broke the Turing test — the race is on for new ways to assess AI. Nature, 619:686–689.