The Emergence of the Shape Bias Results from Communicative Efficiency

Proceedings of the 25th Conference on Computational Natural Language Learning (CoNLL), pages 607–623 November 10–11, 2021. ©2021 Association for Computational Linguistics 607 The Emergence of the Shape Bias Results from Communicative Efficiency Eva Portelance∗ Stanford University† portelan@stanford.edu Michael C. Frank Stanford University mcfrank@stanford.edu Dan Jurafsky Stanford University jurafsky@stanford.edu Alessandro Sordoni Microsoft Research Montreal alsordon@microsoft.com Romain Laroche Microsoft Research Montreal romain.laroche@microsoft.com Abstract By the age of two, children tend to assume that new word categories are based on ob- jects’ shape, rather than their color or tex- ture; this assumption is called the shape bias. They are thought to learn this bias by observ- ing that their caregiver’s language is biased to- wards shape based categories. This presents a chicken and egg problem: if the shape bias must be present in the language in order for children to learn it, how did it arise in lan- guage in the first place? In this paper, we pro- pose that communicative efficiency explains both how the shape bias emerged and why it persists across generations. We model this process with neural emergent language agents that learn to communicate about raw pixelated images. First, we show that the shape bias emerges as a result of efficient communica- tion strategies employed by agents. Second, we show that pressure brought on by commu- nicative need is also necessary for it to per- sist across generations; simply having a shape bias in an agent’s input language is insufficient. These results suggest that, over and above the operation of other learning strategies, the shape bias in human learners may emerge and be sustained by communicative pressures. 1 Introduction Learning new words involves inferring what the ref- erents are in addition to learning a mapping from these inferred referents to linguistic conventions. This is a difficult problem, famously illustrated by Quine’s (1960) problem of referential inscrutabil- ity. Yet, human children manage just fine. A com- mon explanation is that they must be using addi- tional sources of information to learn new words. ∗Corresponding author. †This project was started during an internship at Mi- crosoft Research Montreal. Figure 1: An example experimental trial on novel word learning. Humans typically choose the right picture. One such source is inductive word learning biases, such as the shape bias, which appears in children around two years of age (Heibeck and Markman, 1987; Landau et al., 1988). This bias can be ob- served when children and adults are presented with a novel object and word. They will often assume that this word refers to a class categorized by the shape of the novel object. Thus, when they are presented with more exemplars they will tend to classify those that share the same shape in the same lexical category, rather than those which share the same color or texture, (Figure 1). In this paper, we endeavor to answer two ques- tions: how did the shape bias emerge in our lexical categorization systems? and why does it persist across generations, languages, and cultures? Us- ing neural …

Content truncated to 500 words