Investigating Cross-Linguistic Adjective Ordering Tendencies with a Latent-Variable Model

Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, pages 4016–4028, November 16–20, 2020. c⃝2020 Association for Computational Linguistics 4016 Investigating Cross-Linguistic Adjective Ordering Tendencies with a Latent-Variable Model Jun Yen Leung Guy Emerson Ryan Cotterell , University of Cambridge ETH Z¨urich junyenle@gmail.com, gete2@cam.ac.uk, ryan.cotterell@inf.ethz.ch Abstract Across languages, multiple consecutive adjec- tives modifying a noun (e.g. “the big red dog”) follow certain unmarked ordering rules. While explanatory accounts have been put forward, much of the work done in this area has re- lied primarily on the intuitive judgment of na- tive speakers, rather than on corpus data. We present the first purely corpus-driven model of multi-lingual adjective ordering in the form of a latent-variable model that can accurately or- der adjectives across 24 different languages, even when the training and testing languages are different. We utilize this novel statistical model to provide strong converging evidence for the existence of universal, cross-linguistic, hierarchical adjective ordering tendencies. 1 Introduction Most native speakers of a language would agree that certain adjective orderings are preferable to others. For instance, in English, “the big red dog” sounds natural while “the red big dog” sounds very awkward. Similar ordering preferences have been found to apply universally across the languages in the world: for example, the adjective for “big” in most languages tends to be farther away from the noun, syntactically, than “red.” For an overview of these phenomena, see Cinque (2010). There are many explanatory accounts of cross- linguistic adjective ordering in the linguistics lit- erature, the most popular being hierarchical ten- dencies based on semantic categories of adjectives (Dixon, 1982; Sproat and Shih, 1991; Cinque, 1994, 2010). For instance, Sproat and Shih (1991) and Cinque (2010) note that adjectives describing SIZE tend to be placed further from the noun than those describing COLOR in most languages. However, most of these studies have relied primarily on the judgment of native speakers rather than on cor- pus data, and those corpus-based models that do exist have focused exclusively on English (Shaw and Hatzivassiloglou, 1999; Malouf, 2000; Wulff, 2003; Mitchell, 2009; Dunlop et al., 2010; Mitchell et al., 2011; Hill, 2012; Scontras et al., 2017; Hahn et al., 2018; Futrell et al., 2020). In this paper, we make use of tools and techniques from statistical modeling to provide strong converging evidence supporting a hierarchical theory of cross-linguistic adjective ordering. Specifically, we present a novel interpretable, multi-lingual, latent-variable model of adjective ordering that directly enforces a hierarchy of se- mantic classes and is trained entirely using cor- pus data. We empirically show that our model accurately orders adjectives across 24 different lan- guages, even when tested on languages that it has not been trained on. In doing so, we demonstrate the existence of universal, cross-linguistic, hierar- chical tendencies in adjective ordering. 2 Adjective Ordering Consider the following English phrases, taken from Teodorescu (2006): (1) A beautiful small black purse (2) a. # A beautiful black small purse1 b. # A small beautiful black purse c. # …

Content truncated to 500 words