Chaining and the growth of linguistic categories

Abstract

We explore how linguistic categories extend over time as novel items are assigned to existing categories. As a case study we consider how Chinese numeral classifiers were extended to emerging nouns over the past half century. Numeral classifiers are common in East and Southeast Asian languages, and are prominent in the cognitive linguistics literature as examples of radial categories. Each member of a radial category is linked to a central prototype, and this view of categorization therefore contrasts with exemplar-based accounts that deny the existence of category prototypes. We explore these competing views by evaluating computational models of category growth that draw on existing psychological models of categorization. We find that an exemplar-based approach closely related to the Generalized Context Model provides the best account of our data. Our work suggests that numeral classifiers and other categories previously described as radial categories may be better understood as exemplar-based categories, and thereby strengthens the connection between cognitive linguistics and psychological models of categorization.

Introduction

Language users routinely face the challenge of categorizing novel items. Over the past few decades, items such as emojis, blogs and drones have entered our lives and we have found ways to talk about them. Sometimes we create new categories for novel items, but in many cases we assign them to existing categories. Here we present a computational analysis of the cognitive process by which categories extend in meaning over time.

Lakoff and other scholars (Bybee, Perkins, & Pagliuca, 1994; Geeraerts, 1997; Lakoff, 1987; Malt, Sloman, Gennari, Shi, & Wang, 1999) have suggested that linguistic categories grow over time through chaining, a process that links novel items to existing items that are semantically similar, hence forming chain-like structures of meaning (Lakoff, 1987). Although Lakoff briefly suggests how chaining applies to semantic categories (e.g. the concept of “climbing”), his two most prominent examples of chaining involve grammatical categories. The first example is the classifier system of Dyirbal (an Australian Aboriginal language), which groups together nouns that may not seem closely related on the surface. For instance, the word balan may precede nouns related to women, fire and dangerous things. The second example is the Japanese classifier hon, which can be applied to a variety of long thin objects such as pencils, sticks and trees. Where an English speaker might say “one pencil,” a Japanese speaker must insert the appropriate classifier (here hon) between the numeral and the noun. Although hon is most typically applied to long thin objects, it can also be applied to martial arts contests using swords (which are long thin objects), and to medical injections (which are carried out using long, thin needles). Martial arts contests and medical injections have little in common, but both can be connected to central members of the hon category through a process of chaining.

In Lakoff's work the notion of chaining is coupled with the notion of centrality, which proposes that a category is organized around a central core. Combining chaining with centrality leads to the notion of a radial category, or one that can be characterized as a network of chains that radiate out from a center (Lakoff, 1987; Lewandowska-Tomaszczyk, 2007). Subsequent work in cognitive linguistics relaxes the idea of a single center and allows that radial categories may have “several centers of comparable importance” (Palmer and Woodman, 1999, p 230), but is still committed to the idea that some members of a radial category are privileged by virtue of their centrality. In principle, however, the notions of chaining and centrality can be decoupled. Consider, for example, a category that is constructed by starting with one element and repeatedly adding a new element that is similar to a randomly chosen member of the category. This generative process seems consistent with the notion of chaining, but the categories it produces may take the form of sprawling networks rather than collections of chains radiating out from a center.

Many discussions of chaining within cognitive linguistics are heavily influenced by Rosch and her prototype theory of categorization (e.g., Geeraerts, 1997), but this literature has been largely separate from the psychological literature on computational models of categorization (Chandler, 2017; Polzenhagen & Xia, 2014). The modeling literature includes many comparisons between exemplar models and prototype models of categorization, and the question of whether categories have a central core lies at the heart of the difference between the two approaches. Exemplar models proposes that the representation of a category is no more than an enumeration of all members of the category, but prototype models propose that category representations incorporate some additional element such as a prototype, a central tendency or a set of core examples.¹ Decoupling chaining from centrality means that the process of chaining is potentially compatible with both prototype-based and exemplar-based accounts of categorization, and opens up the possibility of formal accounts of chaining that build on exemplar models like the Generalized Context Model (GCM, Nosofsky, 1986) that have achieved notable success as psychological models of categorization. Here we evaluate a suite of formal models, including a prototype model and a family of exemplar models, and find that an exemplar model closely related to the GCM provides the best account of category growth over time. Our results are broadly consistent with previous work on computational models of categorization, which often finds that exemplar theory outperforms prototype theory when instances of the two are put to the test.

Following Lakoff we focus on grammatical categories, and as a case study we consider how Chinese numeral classifiers have been applied to novel nouns over the past fifty years. As for Japanese classifiers, Chinese classifiers are obligatory when a noun is paired with a numeral, e.g., one [classifier_x] person or two [classifier_y] documents. Although we focus on Chinese classifiers, numeral classifiers are found in many other languages around the world, and have been extensively studied by cognitive psychologists, linguists, and anthropologists (Aikhenvald, 2000; Allan, 1977; Berlin & Romney, 1964; Dixon, Dixon, & Dixon, 1972; Lakoff, 1987). For instance, Allan (1977) has suggested that classifiers across languages often capture perceptual properties such as shape and size, and Aikhenvald (2000) has suggested that classifiers also capture more abstract features such as animacy. Although previous scholars have explored how people assign classifiers to nouns (Gao & Malt, 2009; Tai, 1994), most of this work has not been computational. Our approach goes beyond the small amount of existing computational work (Guo & Zhong, 2005; Morgado da Costa, Bond, & Gao, 2016; Peinelt, Liakata, & Hsieh, 2017; Wen, Gao, & Bond, 2012; Zhan & Levy, 2018) by analyzing historical data and focusing on the application of classifiers to novel nouns.

There are at least three reasons why numeral classifiers provide a natural venue for testing computational theories of category extension. First, they connect with classic examples such as Lakoff's analysis of hon that are central to the cognitive linguistics literature on chaining and category extension. Second, classifiers are applied to nouns, which form a broad and constantly-expanding part of the lexicon, and therefore offer many opportunities to explore how linguistic categories are applied to novel items. Third, the item classified by a term like hon is typically the noun phrase that directly follows the classifier, which makes it relatively simple to extract category members from a historical corpus (e.g., via part-of-speech tags).

Our work goes beyond Lakoff's treatment of classifiers in three important ways. First, we present a computational framework that allows us to evaluate precise hypotheses about the mechanism responsible for chaining. Second, we test these hypotheses broadly by analyzing a large set of classifiers and their usage in natural contexts, instead of considering a handful of isolated examples. Third, as mentioned already our space of models includes exemplar-based approaches that have not been explored in depth by previous computational accounts of chaining. Previous scholars have given exemplar-based accounts of several aspects of language including phonetics, phonology, morphology, word senses, and constructions (Bybee, 2013; Chandler, 2017; Keuleers, 2008; Pierrehumbert, 2001; Ramsey, 2017; Skousen, 1989), and our approach builds on and contributes to this tradition.

Our approach also builds on recent computational work that explores formal models of chaining in the historical emergence of word meanings. In particular, Ramiro, Srinivasan, Malt, and Xu (2018) demonstrated that neighborhood-based chaining algorithms can recapitulate the emerging order of word senses recorded in the history of English. This work found that the best-performing algorithm was a nearest-neighbor model that extends the semantic range of a word by connecting closely related senses. Two earlier studies report that the same nearest-neighbor model also accounts for container naming across languages (Sloman, Malt, & Fridman, 2001; Xu, Regier, & Malt, 2016). This paper compares a suite of models including the nearest-neighbor model that was successful in previous work. We find that our historical data on the growth of Chinese classifiers is best explained by a model that adjusts the nearest-neighbor approach in several ways that are consistent with the GCM (Nosofsky, 1986), an influential exemplar-based model of categorization. Our results therefore suggest that the same categorization mechanisms studied in lab-based tests of the GCM may help to explain how real-world linguistic categories extend over time.