Our research, titled “Linguistic Alignment in Referential Communication: Automatic Detection and Analysis of Its Impact on Shared Conceptualization,” has been accepted for presentation in a poster session at the upcoming workshop “(Mis)alignment in alignment research: a multidisciplinary workshop on alignment in interaction.” The workshop is scheduled for October 19th and 20th, 2023, and will be hosted by the Laboratoire Parole et Langage in Aix-en-Provence, France.
Abstract: Dialogue participants often exhibit considerable similarities in their conceptualizations post-interaction compared to pre-interaction. However, the specific role of linguistic alignment in shaping this conceptualization remains unclear. This study aims to provide greater insight into this question. We introduce automated methodologies for detecting and analyzing form-based linguistic alignment in face-to-face dialogues. Our approach employs natural language processing techniques for efficient large-scale data analysis instead of laborious and possibly subjective traditional annotation methods. By leveraging these techniques, we seek to deepen our understanding of how interlocutors employ language to establish shared conceptualizations. In our work, we consider form-based linguistic alignment by detecting instances of shared expressions - sequences of lemmas used by both dialogue participants for a single referent. We evaluate our approach on a dataset of 67 dialogues where participants play a referential game. This six-round game involves one participant (the director) describing a novel object while another (the matcher) identifies it among several candidates, with roles alternating each round. Our findings show a high prevalence of linguistic alignment, with shared expressions found for nearly all objects in at least one round of dialogue. Additionally, we observe that the more frequently a shared expression is used and the more rounds it spans, the more similar a participant’s post-dialogue conceptualization is to that shared expression. Moreover, our study reveals a positive correlation between the similarity in the interlocutors’ post-interaction names and the attributes of shared expressions, such as the maximum count of rounds and turns in which these expressions are employed. These findings illustrate how automated methods can offer more detailed insights into the processes by which interlocutors create shared understanding through language, providing a promising path to explore further the field of alignment in speech and related behaviors.