Quantification of Mutual Understanding in Task-Based Human-Human Interactions

Abstract

This thesis explores the quantification of mutual understanding in task-based interactions by observing the relation between patterns of repetitions and measures of communicative success. Two important characteristics of mutual understanding have to be kept in mind: it cannot be established for certainty and cannot be directly measured. However, signs of understanding can be detected and quantified, based on two elements: (1) the way by which conversational partners achieve understanding is dependent on their communicative behaviour, and (2) dialogues exhibit repetitions, despite the immense number of possibilities to compose sentences with words that are at our disposal. These repetitions of linguistic choices between conversational partners, a process known as alignment, are argued to play an important role in the establishment of a common ground that leads to understanding. The exact dynamic of alignment ? and related phenomena such as synchrony ? is still under debate, which has created a large body of research interested in determining its scope. However, fewer studies have been conducted that systematically examine its relation with communicative success, and even fewer studies do it in an automatic way that does not require human annotations. It is in this perspective that the research presented here compares repetition patterns to different communicative assessment methods, namely task-success scores, presence of high levels of negative/positive cognitive states, and third-party moderator evaluation. Five corpora with a total of 192 dialogues (about 32 hours) are analysed in terms of other-shared and self-shared repetitions, at different levels of linguistic representations and utterance lengths. The main contribution of this thesis is the establishment of the extent to which repetitions ? categorised as happening outside chance variation ? may function as a proxy measure of mutual understanding. Results suggest a higher proportion of other than self -repetitions happening above chance in task-based interactions. While participants in the position of information givers have a higher volume of speech and use longer utterances, information followers repeat the giver and themselves more. Information givers repeating themselves seem to relate to higher task success, even more so when repeating themselves structurally, in particular for women. Furthermore, familiarity emerged as a decisive factor for success. Participants being familiar with each other unsurprisingly achieved better scores whether they exhibited signs of linguistic alignment or not, however, unfamiliar partners seemed to benefit from alignment, in particular at first attempt of a task. In computer-mediated interactions, both other and self repetitions happened in high proportions, and a significant drop in self -repetitions of long utterances was observed in troublesome dialogues; in interactions monitored by a human facilitator, more encouragements were provided where the method detected less alignment and inversely less encouragement when alignment was present. These two findings highlight the potential of (1) detection of problematic communication, (2) indication of the state of an interaction ? mutual understanding taking place or not, of the described method. It was also found that American speakers repeat themselves more than Scottish speakers. However, in both dialects, familiar participants did not need to exhibit alignment to succeed in the task. Finally, divergence ? taken as the opposite behaviour of alignment ? was very seldom exhibited in the task-based corpora analysed. Altogether, the proxy measure of mutual understanding described in this document stress that the research efforts made in this direction have a great potential both for the improvement of dialogue systems and monitoring critical human interactions.