Work

Towards Human-like Conversations: Repair Strategies and Users’ Mental Model in Task-Oriented Dialogue Systems

Year: 2022
Type: dissertation
Abstract: In the past few years, Natural Language Understanding abilities of Conversational Agents (CAs) have significantly improved.However, the interaction between humans and CAs is still far from perfect: people may employ fairly common linguistic strategies in human communication that are still hard to interpret for machines.While it is important to continue developing neural models that can represent language better and better, total perfection is not a reachable nor desirable goal.Even two humans communicating could have difficulties understanding each other.The strength of natural language is not its infallibility, but rather its ability to signal an error to the counterpart, to manifest the incomprehension, in order to initiate a repair and to bring the conversation back on its tracks.Since in human communication it is more efficient to recognize breakdowns in conversations rather than try to understand all input all the time, the present work applies the repair strategies framework in the context of human-machine interaction.The goal is thus twofold: first, to automatically detect the breakdowns in conversations (that is, when a user signals an incomprehension of sort); second, to categorize those breakdowns according to the repair strategy the human is employing.Once the repair strategy has been correctly identified, there is not a single correct way to react to it: depending on the human counterpart, a certain answer may be more appropriate than another.The present work focus on a feature that is of paramount importance in describing the users: their mental model (MM).That is, their internal representation of a CA and their expectations about its functioning.The second part of the thesis focus on the topic of mental model in the context of CAs and how to recognize it in human-machine dialogues.We propose a simple methodology to gain some insights into the person's MM.We then apply such findings to our dataset in order to obtain double-tagged conversations: in the span of the same dialogue, we know what kind of repair strategy was employed as well as the user's MM that employed them.Being aware of these two information, the CA can elaborate a personalized answer.We tested the personalized answer by means of a user study and found that people did indeed prefer the modified answers.5 6 CHAPTER 1. ABSTRACT and to initiate a repair.Let's observe for instance this dialogue between a family and an Amazon Alexa device 1 :: Alexa! (pause) Myanmar is South or South East Asian?Alexa: Sorry, I'm not sure.Son: Myanmar is South or South East Asian?(Alexa beeps) (Family conversation pauses) Mother: Try again.Son (louder): Is Myanmar part of South East Asia?Alexa: I'm not quite sure how to help you with that.Son: Hmmm. . .Mother: Alexa.Who is South East Asia?What country is East Asia?Alexa: There are no UN recognized countries in Eastern Europe.Son: This is getting annoying.Father:(raised voice): Alexa.Define South East Asian countries.Alexa: This might answer your question.There are eleven sovereign states in South East Asia: Brunei, Cambodia,
Institution: University of Turin
Language: English
Open Access status: green
OpenAlex ID: W7217091316
Cites: 167
Cited by:
Related to: 10