Notes on Machine Translation

Although the quality of machine translation has improved over the years, and really it's quite good now, it's not 100% accurate. For this application, a machine can never substitute a human... But, I simply can't afford professional translations, so this is the best I can settle for. Accordingly, I've made the best of this situation as I can.

Over time, I've devised various strategies in order to ensure these machine translations are as accurate as possible. All of the machine translations posted on this site (those done by me, anyway) will follow most if not all of these instructions.

1. Within a plain text editor, I create a (hopefully) perfect transcription of the document, formatted left-to-right in raw Japanese.


2. I take any given paragraph from the transcript and place it into Google Translate & DeepL (machine translation services), and RomajiDesu (a Japanese-English dictionary service).


3. I compare the initial translations between Google and DeepL to see if there are any major differences. I also use RomajiDesu to view the phonetic reading of the text and to get individual translations of each word. Through this first pass of MTL, I can get a rudimentary understanding of what the text might be talking about.


4. After this, I perform a second pass where each sentence is MTL-ed individually, as well as fragments of compound sentences and anything encased in quotation marks. This is because MTL will often omit entire sentences or fragments of them, and it will even paraphrase quotes. Additionally, I once again consult JP-to-EN dictionaries.


5. Using the new knowledge gained from the first and second passes, I return to the full paragraph in DeepL, and use the correction feature to change any sentence/word into what I believe to be most appropriate. If Google Translate, a dictionary, or something else has a superior translation in certain segments, I try to recreate that within DeepL (to the best of my ability, it's a finnicky process), just as placeholders.


6. Next, I take the draft from DeepL and replace the placeholders with the "best" translations, if necessary. This is also the stage where I format the text. Indentations, bolding, italicizing. More importantly, sentence structure, punctuation, spelling (yes, even spelling). Names of certain people, places, items need to be transliterated on an individual basis, because there is no universally adopted method of Japanese transliteration into Latin. I personally use the Hepburn style with macron diacritics on long vowels, but if there's a different, more common and/or notable form of a name, I'll just use that instead.


7. This whole process repeats until the whole document is finished. I will often go back and review earlier paragraphs if I learn new information during the process.

The most important part of using machine translation is to be somewhat knowledgable on the subject I'm researching. I mean this in a generalized sense. If I'm going to be machine translating interviews from game developers, first I need to know how they think. I don't have to be a game developer, and I don't have to know anything about the specific person, company, or video game I'm researching (although if there's any information available in English, I will absolutely read that first). But just by knowing the technological constraints of the time period and the social atmosphere of the time period, how a programmer might have solved a problem, how a pixel artist might have drawn a sprite, what it was like to be an office worker in a foreign country, what was pop culture at the time... That knowledge leads me in the right direction.

Another equally important aspect is to be able to question anything I read. I guess what I mean is that I must be able to use critical thinking when reading. I often ask myself during this process: Who is the author of this text, and are they different from the speaker in the text? Is there more than one speaker? What is the subject of this text? What time period is the text talking about? Something in the past, in the present, in the future? Who was the target audience of this text? What biases might the author and/or speaker possess?

Scanning Text from Images

Some of my sources, such as magazines, are from photocopied scans, which are basically just photographs, which means I can't copy and paste the text from them. I use Google Lens to scan these images and get transcripts of them. So, first I collect the full transcript in Japanese, and only after this do I begin machine translating the document.

Still, Lens has its flaws, which I will explore below, with details on how I mitigate these flaws to the best of my ability. Because of its high unreliablity, I don't use Lens to machine translate anything unless it's a super simple image, and even then I still cross-reference it with other services just in case. The main purpose of Lens is for transcription, and even then those transcripts must be heavily refined.

• Sometimes Lens will misidentify a character in the transcript, and print the wrong one. Because of this, I try my best to scan each line one-by-one to make sure all of the characters are the same. I just have to eyeball it, but that's better than nothing. To get the correct character, I usually just zoom in a bit more on the image and get a "clearer" shot of it so that hopefully Lens can identify it. But if that fails, I'll have to identify the character through writing it. I usually use Google Translate's handwriting identification feature for this. My calligraphy skills for writing East Asian characters are pretty lacking although I have basic knowledge, and it's even worse to do it on a mouse... But this method is usually successful regardless. I then copy the correct character and place it where it belongs in the transcript.


• Oftentimes Lens will miss a character outright as well. I use the same two methods detailed above to recover the missing character.


• Lens is also terrible at recognizing tategaki (vertical writing) in big paragraphs. It's supposed to be able to do it, and most of the time it can, but sometimes it will scan a particular text horizontally no matter what (no doubt due to the original source's formatting). To get around this unfortunate limitation, I scan each line one at a time. Once I've collected transcripts for each line, I begin consolidating these fragmented lines into whole sentences and paragraphs based on the punctuation marks and text indentations in the original document.


• Regardless of whether the text is tategaki or yokogaki (horizontal writing), Lens will oftentimes see a line break as the end of a sentence, effectively splitting a sentence into multiple parts. Usually, this is obvious, but I always use a plain text editor when refining transcripts to ensure there aren't any secret line breaks hiding from me, and if I find them, I mend the sentence again.