In an ongoing large-scale study, we examine knowledge of classifiers in 300 Japanese heritage speakers (HSs) (age: 10-20). Numeral classifiers are a grammatical device that reflects how speakers categorize objects that they count or quantify, and are obligatory whenever quantity is specified. For example, –hon is a Japanese classifier for long, thin objects (e.g., tree, sword, pole). HSs appear to have some difficulties with classifier-noun mappings (Kan, 2019; Ruiting, 2016). However, it is still uncertain to what extent they understand the semantic criteria (i.e., shape and function properties) driving classifier selection. The current study will examine this question by using both real and nonce objects to test Japanese HSs’ comprehension and production of classifiers. The participants will be tested online. They speak various majority languages and have different levels of exposure to Japanese literacy (e.g., children who attend Japanese-bilingual schools, Saturday schools, or no experience of formal schooling). In a picture-identification task, participants hear a sentence such as: “ni–hon arimasu” “There are two-CLASSIFIER”. At the same time, they will see a computer screen with two pictures: one that corresponds to the classifier –hon (e.g., tree) and the other that doesn’t (e.g., cat). The trials will alternate real and nonce objects. In the production task, participants are asked to count the numbers of real or nonce objects depicted on screen. We hypothesize that literacy as well as other proxies for input/exposure quality/quantity to Japanese will predict knowledge of classifiers—especially in the nonce object condition, where the semantic criteria are tested—since such related variables are good candidates for determining the extent heritage speaker bilinguals will differ from monolinguals and from each other (Bayram et al., 2018; Kupisch & Rothman, 2016).