Suppose you hear a word like Korean tuˈgɯndugɯn and two possible meanings ‘heartbeat’ or ‘gentle movement’. Most people pick the first (as we found in 2016). Or suppose you hear that same word and see the meaning ‘heartbeat’ and you’re asked to rate, on a scale of 1 to 5, how well form and meaning fit together. Most people will go high (4.7 on average, as we found).
Okay, so people have intuitions about the fit between forms and meanings — even of words they’ve never encountered before. But can we explain those intuitions? Can we use that explanation to predict what they do in guessing or rating tasks? That is the question we set out to answer in The Anatomy of Iconicity.
Punselie, Stella, Bonnie McLean, and Mark Dingemanse. 2024. ‘The Anatomy of Iconicity: Cumulative Structural Analogies Underlie Objective and Subjective Measures of Iconicity’. Open Mind 8 (September): 1191–1212. doi:10.1162/opmi_a_00162.
Our work is published in open access and we’d be thrilled if you go read the full thing. Here I’ll just summarise it using three figures from the paper.
Structure-mapping
In this article we follow a long line of luminaries who’ve argued that iconicity can be seen as a form of structure-mapping. We pay special credit to Sarah Taub (2001), who made this argument in a classic —and undercited— paper for both visual and speech-based signs. 1 The figure above is an adaptation of two diagrams from her work (reproduced with permission).
Structure-mapping is most closely associated with the work of Dedre Gentner (e.g., Gentner 1983). It revolves around structural correspondences — in the case of iconicity, between aspects of linguistic form and meaning. When we say two fingers can iconically stand for the legs (Figure 1, A), we mean that they share structural properties that can be brought into correspondence. And when we say that a word like ding can iconically stand for the sound of a bell, we mean that both these acoustic events share properties that can be brought into (structural) correspondence (Figure 1, B).
Structure-mapping has a number of interesting features that can help us better understand natural language iconicity. It is selective: its goal is not to create a perfect replica, but to present a subset of salient structural correspondences. To some extent this follows from the fact that structural properties of source and target differ, and so offer different affordances for establishing correspondences. Importantly, the correspondences can be cumulative: a word and its meaning can be tied by multiple layers of structural analogies, and certain form/meaning combinations may offer more such layers than others. Which is what brings us to the main design.
Triangulation
The overall methodological approach can be best explained with the example I started out with: Korean tuˈgɯndugɯn ‘heartbeat’. What makes this word work so well? The proposal we work out is that there are multiple structural correspondences between form and meaning: the high-level repetition captures the iterative nature of the beating heart, the individual syllables map onto what some folks call the ‘lub’ and ‘dub’ of the heartbeat sound, and there is even an acoustic similarity between the speech and the nonlinguistic sound. If we allot one ‘iconicity point’ for each of these, tuˈgɯndugɯn ‘heartbeat’ gets a cumulative iconicity score of 3.
One prediction we make is that words higher in cumulative iconicity will also be rated as more iconic and will be more guessable, all else being equal. This is indeed borne out for tuˈgɯndugɯn (panels B and C). In the paper, we do this for a few hundred words. If structure-mapping indeed underpins iconicity, then word-level cumulative iconicity scores should help us predict people’s intuitions about words and meanings. Then structure-mapping will help us explain people’s intuitions about iconicity.
This is a form of triangulation: using multiple sources of evidence to arrive at a more precise characterization of a particular phenomenon, in this case iconicity.
Vindication
The third figure (probably my favourite) puts the pieces together. It first shows, for all 239 ideophones included in the study, how people rated them for subjective iconicity (panel A). There’s quite a range of variation here: it is by no means the case that every supposedly iconic word “feels” equally iconic to non-users of a language (nor even to users of that language).
The data points are coloured by their cumulative iconicity (lighter = higher in cumulative iconicity), which foreshadows the result seen up close in panel B. Here, we unpeel the cloud of points into layers according to cumulative iconicity. Now we see that iconicity ratings are neatly predicted by cumulative iconicity: the higher an ideophone scores in cumulative iconicity (from 0 to 5), the higher it is rated.
Importantly, this is a one-way street. It is hard for an ideophone to have high cumulative iconicity and not be rated as highly iconic; but there are ideophones that are rated as highly iconic yet don’t have a high cumulative iconicity score in our system. This is because we operationalised cumulative iconicity in a partial and selective way: we capture only 5 broad types of structural correspondences. The ratings data suggests there are likely more, and so our triangulation method also offers a way of identifying ideophones that may yield evidence of iconic assocations not yet covered.
Student-led slow science
This was an incredibly fun paper to work on. It started, for me, way back in 2017 when I devised a first version of a way to code structural correspondences in order to better ground our understanding of iconicity in natural languages. I realized this would be best done as a collaboration and found first author Stella Punselie willing to work on it for her BA. Stella helped pilot and subsequently refine the coding scheme, collected and analysed iconicity rating data, and delivered a stellar BA thesis. Already then it was clear this would probably lead to a paper. In my research lab, this typically means a student-led publication (indeed this is the fourth such paper since 2020).
This one took longer to materialize because a pandemic intervened. We had a near-complete draft in 2020 but then Stella finished an MA in Leiden and found work elsewhere while I shifted focus to a different line of work, so we left it on the back burner. We did get very helpful input from Bonnie McLean, who also contributed some exploratory analyses and made a Shiny app for interactive exploration. Slowly but surely the paper took form. A final impulse came when we realized, due to a bunch of other papers, that it would be very helpful to have this out.
We decided to submit to Open Mind, a premier diamond open access journal published by MIT Press. We considered this cognitive science journal a good fit for this work given that we build on classic cogsci work (structure-mapping) and connect it to natural language iconicity. The paper was pretty mature at submission time, in part because of our unrushed, theory-infused, empirically grounded way of working. It was not some “output” we needed, it was a labour of love and a very nice and satisfying project to work on. We’re very happy with how it’s turned out and hope you’ll give it a read!
References
- Emmorey, K. (2014). Iconicity as structure mapping. Philosophical Transactions of the Royal Society B: Biological Sciences, 369(1651), 20130301. doi: 10.1098/rstb.2013.0301
- Gentner, D. (1983). Structure-mapping: A theoretical framework for analogy. Cognitive Science, 7(2), 155–170. doi: 10.1016/S0364-0213(83)80009-3
- Punselie, Stella, Bonnie McLean, and Mark Dingemanse. 2024. ‘The Anatomy of Iconicity: Cumulative Structural Analogies Underlie Objective and Subjective Measures of Iconicity’. Open Mind 8 (September): 1191–1212. doi:10.1162/opmi_a_00162
- Taub, S. F. (2000). Iconicity in American sign language: Concrete and metaphorical applications. Spatial Cognition and Computation, 2(1), 31–50. doi: 10.1023/A:1011440231221
- Tufvesson, S. (2011). Analogy-making in the Semai Sensory World. The Senses and Society, 6(1), 86–95. doi: 10.2752/174589311X12893982233876
- We also cite many others who have built on this insight, from Gattis 2004 and Emmorey 2014 for sign languages to Tufvesson 2011 and Thompson & Do 2019 for spoken languages.[↩]