Does a Japanese LM Process Kanji and Kana Words Differently? (Hypothesis)

Wang et al. (2025) found that multilingual language models (MLMs) generally perform three distinct steps given a task of recalling factual knowledgeE.g., asking a model to state the capital of a country.. First, an MLM processes the input sentence syntacitcally from the positions to its tokens to the relations each token hold against each other. This is done in the prompt's language for a balanced MLM, and is done in English for a biased oneI.e., a model trained using predominantly English dataset like LLaMA2. The second stage, object extraction, is done in the model's latent language. This depends on the pretraining distribution, but could be heavily biased towards English. Lastly, the model performs target language-specific processing. In this stage, the model transitions to the language of the task, processing semantic features unique to the target language. This finding shows that MLMs have a "shared concept space" that is agnostic to language where "thinking" happens, before returning next tokens in the target language. This poses an interesting question in humans, too. Do multilingual humans process input in an "abstract language" before translating it back to human language? Does this depend on the human's fluency? For example, does a native English speaker with A2-level German translate every German input into English first, produce the answer in thought, and translate it into a German response?

Japanese and Its Three Character Systems

There is another curious question that is specific to the Japanese language. Japanese language regularly uses three types of characters. Kanji, hiragana, and katakana. For simplicity, I would refer to the latter two as simply "kana". Many words have kanji representations, but usually only loanwords are written in katakana. However, regardless of the type of word, every Japanese text comes with furigana, the phonetic information of the text, usually written in hiragana. This means that every word written in kanji can be written in hiragana too, regardless of practical relevancy. Culturally, hiragana provides friendliness, fun, and easy-to-approach nuance. However, it also has a functional use: readability. Hiragana is just like the English alphabet, in a way that a combination of multiple hiragan characters make up a Japanese word in the dictionary⊕As opposed to kanji which generally holds one meaning per character.. This makes hiragana useful for street signs that are meant to be read by everyone, especially children.