Back when trying relative search, I marked the kana characters with roman letters to find sequences, since characters are 4 sprites big, I began the first sequence changing the first tile of the character, and the next sequence changing the second tile and so on.
When the first dialog is shown, the following tiles are loaded into VRAM (at offset 21504):
2U 1D 1H (Kanji) (Kanji) 4G 4I 4M 2z 2f 4R
(2U is the Kana which has an “U” on its second sprite)
Following Vehek’s input, I’m hoping to find a relative sequence for this somewhere in the ROM; what I’m thinking now is that there should be some sort of table which indicates what sprites to load for each dialog, and another one which says what actual sprites to use based on the ones that were loaded already.
If this is true and I manage to find both, I’ll be able to make a program to extract the script.