Examples of Machine Translation

 I thought it would be interesting to see how a machine may translate an excerpt from a well-known source that has been around for quite a while and thus has been translated many times throughout the years. I wanted to see if more more human "input" would impact the accuracy or overall tone at all.

人間失格/No Longer Human by Osamu Dazai (trans. by Donald Keene in 1958)

Original Text:

私は、その男の写真を三葉，見たことがある。一葉は、その男の、幼年時代、とでも言うべきであろうか、歳前後かと推定される頃の写真であって、その子供が大勢の女のひとに取りかこまれ、（それは、その子供の姉たち、妹たち、それから、従姉妹たちかとたちかと想像させる）庭園の池のほとりに、荒い縞の袴をはいて立ち，木を三十度ほど左に傾け、醜く笑っている写真である。

Donald Keene's translation of the very first bit of the book:

    I have seen three pictures of the man.

    The first, a childhood photograph you might call it, shows him about the age of ten, a small boy surrounded by a great many women (his sisters and cousins, no doubt). He stands in brightly checked trousers by the edge of a garden pond. His head is tilted at an angle thirty degrees to the left, and his teeth are bared in an ugly smirk. 

ChatGPT Translation:

    I have seen three photographs of that man.

   One of them is a photograph from what might be called the man’s childhood—or perhaps from around the age of ten, if I had to guess. In it, the child is surrounded by a large number of women (they appear to be his older and younger sisters, and perhaps his cousins), standing beside a pond in a garden. He is wearing a hakama with broad, rough stripes, tilting his head about thirty degrees to the left, and smiling grotesquely.

Google Translate:

    I have seen three photographs of the man. 

    The first dates from what might be called his early childhood—a time when he was presumably around ten years old. In this picture, the child is surrounded by a crowd of women (presumably his older and younger sisters and female cousins); he stands by a garden pond wearing *hakama* trousers with a bold striped pattern, his body tilted about thirty degrees to the left, wearing an ugly, grinning smile.

    

Commentary:

    As I expected, ChatGPT recognized the passage as the beginning of No Longer Human right away. One of the things I found interesting was the word choice the AI used for descriptive language, especially the little boy's clothing. In Donald Keene's translation, he refers to the boy's clothes as trousers, but ChatGPT stays true to the original text in this sense and uses "hakama" instead. The human translation is catering to the average American reader, who would most likely not know what a hakama is. His consideration of the audience's perspective, and ChatGPT's lack thereof, is definitely the most intriguing difference between the two. Keene chose to use "trousers" because he is translating for comprehension by the audience, not purely for accuracy to the original text. Also the AI seems to truly be going through the sentences one by one and translating them, rather than having it flow in a way that sounds natural or easily consumable. It has too many mid-sentence pauses and each additional descriptive detail feels like an afterthought that was just thrown in. 

    The Google Translate translation is pretty clearly less refined than that of ChatGPT, but there are some similarities between the two. Both machine translations come across as way more unsure of certain details than in the translation by Keene. He interprets the passivity in Dazai's prose as more of an assumption the narrator makes, rather than a conscious observation (I'm not sure how else to say this, sorry if it makes zero sense). 

    Overall, the machine translations are accurate, but cling too tight to the original text to properly convey what it's truly saying to an English-speaking audience. They simply relay information, and lack complexity due to a mechanical narration style, unlike Donald Keene's translation.

Madison Blecha