Wednesday, 11th October - 10am Livia Polanyi : Seminar

Title: Modeling Topic Shifting in Conversational Interaction: A Step Towards More Human Digital Companions

Abstract: 

Developing computational systems that can be true digital companions not merely digital assistants requires an understanding of how unremarkable, “normal” conversational interaction takes place and in particular, how human interactants shift the topic of conversation from talking about one topic to either gradually or abruptly talking about something else entirely. In this talk, after first summarizing the constraints associated with appropriate conversational behavior, I will introduce a dynamic feature based notion of conversational topic that includes both linguistic and interactional information and changes with every utterance. I will then argue that to be seen as an appropriate next utterance, that next utterance must be Informative, Coherent and Relevant to the interactional context. In the case of step-by-step topic shift, a Coherent next utterances will share at least some linguistic features with an utterance located on the right edge of a hierarchical parse tree of the discourse. While traditional Relevance Theory (Sperber and Wilson, 1995) can account for the Relevance of gradual topic shift understanding abrupt, disjunctive topic shift requires an entirely new approach to Relevance to explain why some information is appropriate to introduce in a specific conversation among specific individuals while another, apparently similar piece of information, may not be. To account for this phenomenon, I will introduce a new multi-dimensional model of Conversational Relevance based on the notion “The Closer to Me the More Relevant” that makes explicit what is relevant to human interactants. Based on this information, a Relevance Score (RS) can be assigned to an utterance deployed to introduce a new topic. If the RS equals or exceeds some predetermined minimal score , the utterance is appropriate, and the topic shift successful. In conclusion, because in natural conversation the somewhat risky interactional work of introducing new topics is a task shared among all the people conversing with each other, I will argue that an implemented computational model based on the Conversational Relevance Score would enable Conversational AI Agents to interact much more like Digital Companions than today’s Chatbots and Digital Assistants. Building those new Agents will be left as an exercise for interested researchers.

References:

“Harvey Sack’s Lecture Notes, Lecture 5, Spring 1972” quoted in Gail Jefferson. 1984. “On Stepwise transition from talk about trouble in Atkinson and Heritage (eds), Structures of social action: studies in conversation analysis. Cambridge University.

Sperber, Daniel and Deidre Wilson. 1995. Relevance: Communication and Cognition. Oxford: Wiley Blackwell.

Bio:  

Throughout a long career both within academia (University of Amsterdam, Rice University, Stanford University [Honorary Professor]) and outside (BBN Labs, FujiXerox Palo,Alto Laboratory, Powerset (NLP Search Startup, Microsoft).I have worked on many aspects of the structure, interpretation and use of discourse in fields including Theoretical, Computational and Socio- Linguistics, AI, Literary Theory, Cultural Anthropology and Economics. Beginning with my work on everyday narrative published in Telling the American Story:Linguistic, Social and Cultural Constraints on Stories in Conversation (MIT Press 1985) and continuing through work with Remko Scha and Martin van den Berg developing and later overseeing implementation of The Linguistic Discourse Model which introduced the syntactic framework of Open Right Trees that together with DRT formed the basis of S-DRT, I have been concerned with exploring how language above the sentence operates in natural interaction. A few other detours have had me straying into Sentiment Analysis (with Annie Zaenen), Knowledge Management and the role of cultural differences in remote working teams. My most recent work concerns Topic Shift in Conversation, sparked in part by current interest in Conversational AI which has not yet addressed important dimensions of human-human interaction that must be modeled to make the jump from Conversational Assistants to Conversational Partners.

 vCal  iCal