ChatGPT is sufficiently well known to warrant critical scrutiny, and for this project we wrote a paper, developed a website where we track open-source instruction-tuned large language models, designed a poster for presentation at the ACM conference on Conversational User Interfaces (CUI’23) and, yes, even designed a logo that combines a key image of the open source movement with a variation on ChatGPT’s corporate logo.
Liesenfeld, A., Lopez, A., & Dingemanse, M. (2023). Opening up ChatGPT: tracking openness, transparency, and accountability in instruction-tuned text generators. ACM Conference on Conversational User Interfaces (CUI ’23), July 19-21, Eindhoven. doi: 10.1145/3571884.3604316PDF
A. Overview of included languages with dataset size in hours and top 3 sequentially identified response tokens as transcribed in the corpus. B. Location of largest speech community. C. Assessing the impact of sparse data on UMAP projections using three samples of Dutch response tokens. A look at the full dataset (a) and random-sampled subsets of decreasing size (b, c) suggests isomorphism across scales and interpretability of clustering solutions as small as 150 tokens.
Liesenfeld, A., & Dingemanse, M. (2022). Bottom-up discovery of structure and variation in response tokens (‘backchannels’) across diverse languages. Proceedings of Interspeech 2022, 1126–1130. doi: 10.21437/Interspeech.2022-11288PDF
The relationship between the two parts of a behavior pair can vary on five dimensions, as outlined in this table. For each dimension, we visualize two different relationships between instances of behavior—one with a solid arrow and one with a dashed arrow. For meaning, we use tangram figures to visualize the referent of speech and/or gestures
Rasenberg, M., Özyürek, A., & Dingemanse, M. (2020). Alignment in Multimodal Interaction: An Integrative Framework. Cognitive Science, 44(11). doi: 10.1111/cogs.12911PDF
Shooing words —words that people use to chase away chickens— turn out to be highly similar across unrelated languages. These illustrations by Josje van Koppen accompanied a write-up about my serendipitous finding in popular science magazine Onze Taal.
The actual table from my paper looks a lot less exciting, but it does contain additional information about language families and about words for ‘chicken’ in the same set of languages. The basic conclusions is that words for ‘shoo’, but not ‘chicken’, show strong convergence towards sibilant sounds in 17 languages from 11 unrelated language families.
Illustrations from: Renckens, Erica. “‘Ksst!’ Het Lokken En Wegjagen van Dieren.” Onze Taal, 2020.
Interactional challenges to be negotiated in recruitment sequences, along with some of the interactional practices mobilized to address them.
Not a figure, I know, but sometimes tables are the only way to bring multidimensional problem space into view. In this case, the table is also a map to the resources discussed in the paper.
Dingemanse, M. (2020). Recruiting assistance and collaboration: a West-African corpus study. In S. Floyd, G. Rossi, & N. J. Enfield (Eds.), Getting others to do things: A pragmatic typology of recruitments (pp. 369–421). Language Science Press. https://doi.org/10.5281/zenodo.4018387PDF
The canonical syntactic home of ideophones in Siwu is toward the end of the clause. A finer analysis of patterns of occurrence in the corpus reveals a number of constructions in which ideophones can occur. The five most common constructions, together accounting for 95 % of ideophone tokens, are shown here.
This type of visualization —a table with horizontal bar plot— has no name as of yet. It uses the same logic as E.J. Tufte’s sparklines, which also display numerical information inline.
Dingemanse, M. (2017). Expressiveness and system integration. On the typology of ideophones, with special reference to Siwu. STUF - Language Typology and Universals, 70(2), 363–384. doi: 10.1515/stuf-2017-0018PDF
Using elementary properties of interactional resources, we can capture commonalities and differences between repair formats in principled and precise ways. For instance, to capture the distinctions between four repair initiation formats in English (as presented in Sidnell 2010), we can use the following three properties: Question (is there a content question word?), Repetition (does the repair initiator repeat some material from the prior turn?) and Confirmation (does the repair initiator make confirmation relevant in next turn?).
Dingemanse, M., & Enfield, N. J. (2015). Other-initiated repair across languages: towards a typology of conversational structures. Open Linguistics, 1, 98–118. doi: 10.2478/opli-2014-0007PDF
Herb Clark, building on Austin’s (1962) distinctions of levels of speech acts, notes that successful communication is grounded in joint actions by speaker and addressee at at least four distinct levels. In the Austin/Clark action ladder, higher levels depend on lower levels in terms of causality (higher levels are implemented by means of lower ones) and entailment (completion of a higher level entails completion of the ones below it). As a corollary, the action ladder exhibits the property of “downward evidence”: evidence that B recognized A’s intended action (level 4) is also evidence that B succeeded in interpreting A’s words (level 3), that B correctly identified the words (level 2), and that B attended to A’s vocalisation (level 1). All four levels are involved in building mutual understanding, and each of them can be a locus of trouble.
Dingemanse, M., Blythe, J., & Dirksmeyer, T. (2014). Formats for other-initiation of repair across languages: An exercise in pragmatic typology. Studies in Language, 38(1), 5–43. doi: 10.1075/sl.38.1.01dinPDF
