Synchrony in Human-Bonobo Dialog Carl Vogel * , Maria Koutsombogera * , Justine Reverdy * , Anna Esposito † * Computational Linguistics Group, Trinity College Dublin, The University of Dublin, Dublin 2, Ireland Email: {vogel,koutsomm,reverdyj}@scss.tcd.ie † Universit` a degli Studi della Campania “Luigi Vanvitelli”, Caserta, Italy and International Institute for Advanced Scientific Studies (IIASS), Vietri sul Mare, Italy Email: iiass.annaesp@tin.it Abstract—Past research has suggested that bonobos have demonstrated capabilities of successful linguistic interaction with humans. A transcript of interactions linked to such claims is analyzed using methods that have been deployed to assess mutual understanding in human communication. Evidence of dialog plan maintenance on the part of the bonobo is visible, but not of understanding of the human dialog partners (nor vice versa, mostly). I. I NTRODUCTION Successful interspecies communication between humans and other creatures appears to be possible, to the extent that researchers in cognitive infocommunications 1 have seen ways to improve technology interfaces for human-human communi- cation on the basis of human-dog communication [4]. Others seek to improve communication between humans and non- humans [5]–[7]. Language may be understood as a technology that is adapted for communication [8], [9]. The extent to which interspecies communication involving human language has been successful is an open question. A published analysis of human-bonobo communication [10] focuses on the role of repetition in establishing cohesion in communication, towards a general argument attributing linguistic prowess to the relevant bonobo, Panbanisha: “we wish to argue that the bonobo’s discursive contributions are non-random and that this interspecies communication takes the form of a cohesive conversation with active co-construction by both parties” (p. 23). We take this assessment to be based on qualitative interpretation. The present work re-analyzes the data reported there, quantifying the extent to which the linguistic repetition behavior significantly diverges from what might be expected in a random process. The method of analysis builds on past analyses of conversations focused on assessment of synchrony and engagement among participants [11]–[15]. We build on past analysis of the bonobo data [10], [11]. Application of the method of analysis advocated here leads to the conclusion that there is insufficient support for the claim quoted above to the effect that Panbanisha’s conversational engagement was non-random. Repetition provides an fallible index of interlocutor en- gagement in conversation. Repetition has many functions in conversation beyond signalling synchrony [16]. Further, the illusion of interlocutor participation may be sustained by re- peating elements of discourse without attending to the content, and indeed, one way to fake engagement in conversation is to repeat the odd word or phrase uttered by an interlocutor. For 1 See [1]–[3]. this to be a successful approach, it is necessary that measures of lexical and sub-lexical repetition in un-natural conversations are significantly different than in natural conversational interac- tions. A conversation in which interlocutors talk to each other about unlinked topics would be un-natural, and any lexical overlap would be random. Nonetheless, in natural dialog, it may happen that repetitions are distributed randomly. The point has been made that language is not a random process [17]. When one considers distributional properties of natural language, in particular the facts of the frequency relations between successive pairs of items in rank-frequency orderings of words used, one would probably imagine that repetition in natural language is inevitable: it is hardly a surprise for successive statements by distinct speakers to contain the word “the”. It becomes interesting to consider how much and what kind of repetition in conversation is interesting and revealing of enagement and mutual understand- ing. Recent work has attempted to demonstrate the extent to which repetition in dialog correlates with task success [18]– [20]. Other work is developing profiles of dialogs according to levels of repetition effects (repetition of self, and repetition of others) as measures of interlocutor synchronization and relative involvement in dialog, in support of quantifying evidence of mutual understanding [11]–[15]. We use a Monte Carlo approach in which natural dialog transcripts are compared with randomized versions of the turns. Using past categorization of such approaches [21], the method constructs “full-sample” permutations. Counts of linguistic event types in the natural dialog are compared with counts of the same types in the randomizations. To obtain the randomiza- tions the contributions of each participant are processed into a sequential data-structure, with one cell for each turn. The data- structure is then re-indexed using a random process without replacement, and this re-indexed structure is used to write revised transcripts in which the same turns ultimately happen, but nearly certainly in a different order. Measurements made for the actual dialog are compared with average measures made on the re-ordered dialogs, running the re-ordering process ten times. Here we focus on lexical and lexical sequence repetition counts in successive turns – any turn has a speaker: we consider the count of items shared between that turn and that speaker’s last turn (self-repetition) and the counts shared between that turn and each other speaker’s last turn (other-repetition). In principle, temporal overlap may also be considered as a metric of engagement; however, this data is not available in the bonobo-human data addressed. §II describes the data related to claims of interspecies