Bordag, Gor, and Opitz (Reference Bordag, Gor and Opitz2021) present the Ontogenesis Model of L2 Lexical Representation (OM) as an explanation of how words in a second language are represented in memory. According to the model, a word is represented by six quantities: three indicating how well its orthography, phonology, and semantics are learned and three representing how strongly those representations are interconnected. Over time, a word's orthographic, phonologic, and semantic representations become better learned and inter-connected (see Figure 4 of the article). As outlined in the paper, that framework provides insight into a broad range of lexical behaviours and empirical regularities from the study of second-language acquisition.

The OM is a timely and valuable contribution. It presents a broad and integrative approach to think about lexical representation. It confronts complexities of multilingual representation. It acknowledges the necessity of developing formal models to handle the scale of a problem as large as language. However, the OM stops short of its principal goal.

Although the OM is a theory of lexical representation, it does not define what a lexical representation is or how a lexical representation develops as a function of language experience. Rather, it assumes words have lexical representations; and focuses, instead, on modelling how quickly and how completely those representations are learned. Thus, if one asks what the OM knows about <CAT>, it can report how well <CAT> is learned but cannot report what <CAT> means. In that sense, the OM is more like a theory of associative learning than a theory of mental representation.

As a point of comparison, consider current models of lexical representation. Theories such as HAL (Lund & Burgess, Reference Lund and Burgess1996), LSA (Landauer & Dumais, Reference Landauer and Dumais1997), BEAGLE (Jones & Mewhort, Reference Jones and Mewhort2007), Word2Vec (Mikolov, Sutskever, Chen, Corrado & Dean, Reference Mikolov, Sutskever, Chen, Corrado and Dean2013), and GloVe (Pennington, Socher & Manning, Reference Pennington, Socher and Manning2014) are fully specified models that, when applied to a record of natural language (e.g., newspapers, novels, or internet chatter), derive a unique vector to represent each word. Once derived, those vectors can be decomposed into their representational elements (Hollis & Westbury, Reference Hollis and Westbury2016), applied, and compared to human lexical behaviour. Consistent with the OM's goals, those theories track changes in lexical behaviour as a function of age (Montag, Jones & Smith, Reference Montag, Jones and Smith2015), reading history (Aujla, Reference Aujlain press), culture (Johns & Jamieson, Reference Johns and Jamieson2019), and multilingual language exposure (Johns, Sheppard, Jones & Taler, Reference Johns, Sheppard, Jones and Taler2016). Those representations can also stand in for human knowledge in computer models of memory (Johns, Jones & Mewhort, Reference Johns, Jones and Mewhort2021), decision making (Bhatia, Reference Bhatia2017), language production (Johns, Jamieson, Crump, Jones & Mewhort, Reference Johns, Jamieson, Crump, Jones and Mewhort2020), and cognitive disorder (Johns et al., Reference Johns, Taler, Pisoni, Farlow, Hake, Kareken, Unverzagt and Jones2018). Thus, whereas the OM provides a framework for thinking about lexical representations, modern theories already articulate methods to directly and computationally derive those representations.

In summary, the OM is an intriguing theoretical framework to explain second-language learning. However, its breadth of vision obscures its specificity of explanation. Our goal is to encourage a formalization of the OM, perhaps grounded in a theory we have named or in a different kind of theory that we have not named (Griffiths, Steyvers & Tenenbaum, Reference Griffiths, Steyvers and Tenenbaum2007; Jamieson, Avery, Johns & Jones, Reference Jamieson, Avery, Johns and Jones2018; Kwantes, Reference Kwantes2005). No matter the outcome, a formal expression of the OM will allow researchers to interrogate the model, compare it to state-of-the-art language models, and leverage its insights to make discoveries. We are hopeful that future work can meet that goal.