Woman or tennis player? Visual typicality and lexical frequency affect variation in object naming

Abstract

Speakers often use different names to refer to the same entity (e.g., “woman” vs. “tennis player”). We here explore factors that affect naming variation for visually presented objects. We analyze a large dataset of object names with realistic images and focus on two factors: visual typicality (of both objects and the contexts they appear in) and name frequency. We develop a novel computational approach to estimate visual typicality, using image representations from Computer Vision models. Specifically, we compute visual typicality as similarity between the representation of an object/context to the average representation of other objects/contexts of its nominal class. In contrast to previous studies, we not only study the name used by most annotators for a given object (top name), but also the second most frequently used (alternative name). Our results show that the top name and the alternative name pull in opposite directions. People’s naming choices are more varied for objects that are less typical for their top name, or more typical for their alternative name. They are also more varied when the top name has relatively low frequency (for alternative names, the opposite effect may be present but the data are not conclusive). Context typicality instead does not show a general effect in our analysis. Overall, our results show that visual and lexical characteristics relating to name candidates beyond the top name are informative for predicting variability in object naming. On a methodological level, we demonstrate the potential of using large scale datasets with realistic images in conjunction with computational methods to inform models of human object naming.

Effects of task and visual context on referring expressions using natural scenes

“…However, there may be a general bias towards the visually more salient object (here the batter in the front), which speakers take into account when choosing their expression. Similar considerations can be made for visual typicality with a bias towards the more typical candidate object for a given name (for an analysis of typicality effects on lexical choices see Gualdoni, Brochhagen, Mädebach, & Boleda, 2022).…”

Lexical choice and word formation in a taboo game paradigm

“…Similarly, Graf et al (2016) asked participants to name one of three presented images, which could be of the same or different basic-level or superlevel category, and investigated as their dependent variable the level of description of the produced labels. In very recent studies, Gualdoni and colleagues (Gualdoni et al, 2022(Gualdoni et al, , 2023 examined the ManyNames dataset Silberer, Zarrieß, Westera, & Boleda, 2020), which contains different human-generated image labels for a large set of images -in the vast majority of LEXICAL CHOICE IN A TABOO GAME PARADIGM 5 cases multiple labels with different production frequencies per image. Using computer vision models, Gualdoni and colleagues demonstrated that this naming variation could be predicted from the image typicality (see also ) for these different labels (Gualdoni et al, 2022), and developed a computational measure of label informativeness that predicts the level of description of the produced labels (Gualdoni et al, 2023).…”

From language development to language evolution: A unified view of human lexical creativity

“…We operationalized visual similarity using computationally derived visual representations of meanings. We followed a two-step proce-dure, drawing on existing work (15,31). First, we used a computer vision model (32) to produce representations for images of instances of meanings (e.g., for images of dogs for the meaning "dog").…”