I was talking to my wife about how we did Embedded Hiragana and Katakana yesterday, and she thought it would be fun to teach romaji, so I made one.
It is a multilayer perceptron with 4 alphabetic characters as input (embedded in vector and 16 dimensional), 256 and 128 dimensional intermediate layers, and 37 dimensional output with 10 dimensional*3 embedded vector of kana characters, with OneHot for output length and input consumption.
When they don’t learn well, it’s usually the teacher’s fault.
nishio → nikio
- Nice try Nikio… I hope you at least answer what I taught you…
- The learning data included only Kunrei-style romaji, so “si→shi” was taught, but “shi→shi” was not.
ni → mi shi → ki si → ki o → Pi nisio → nisipi
- You’re starting to sound like a high school girl.
- I wondered why “mikipi” when used individually becomes “nisipi” when connected, and found that the way the training data was created was wrong.
- Initially, the training data was entered as “nish” and the first two letters of “ni” were added randomly after “ni” so that “ni” was returned.
- On the other hand, when testing, if the given string is less than 4 characters long, it is marked “ni$$” with a sentence end marker after it.
- Since "" is seen at the end of a sentence in the test, it causes confusion.
- As a result, the nisipi
- I’ve also mixed $ into the training data and now I can read nishio properly.
chopin → Chipin
- I thought it was odd that you could read nishio but not read this as chopin, so I did some research.
-
chi → chi
-
cha → chichi
-
chu → chichi
-
cho → chichi
- In the guess result output section, when outputting more than one kana character at a time, I mistakenly repeated the first character, which was an implementation error on my part.
-
Check if all questions are answered correctly. - > du → du (A: zu) - > du → deux (A: deux) - They were bullying the students by asking the same questions with different answers they were looking for. - Not sure if it’s a “zu” or a “de,” so it’s interesting to mix them up and make it a “de.”
This page is auto-translated from /nishio/ニューラルネットでローマ字を読む using DeepL. If you looks something interesting but the auto-translated English is not good enough to understand it, feel free to let me know at @nishio_en. I’m very happy to spread my thought to non-Japanese readers.