Machine Translation - Sloane

Sloane McLean 

For this experiment, I chose a story called 「夢十夜」by 夏目漱石. I used both ChatGPT and Google Gemini to get two separate translations for discussion.

Firstly, I tried this sentence: 「腕組をして枕元に 坐っていると、 仰 向に寝た女が、静かな声でもう死にますと云う。」

ChatGPT's translation: "As I sat by the pillow with my arms folded, the woman lying on her back said in a quiet voice, ‘I am going to die now.’"

Google Gemini's translation: "When I sat by her pillow with my arms crossed, the woman lying on her back quietly said, "I'm going to die now."

To me, both of these translations sound a bit awkward, with Gemini's being decidedly more so. I think that while the wording is pretty accurate in both versions, Gemini makes the already lengthy sentence sound borderline mechanical. It lists the state of the narrator (crossing their arms) then the state of the woman (lying on her back), then tells us her action (announcing her imminent death). It feels like every aspect of the sentence is turned into an action, listing out what happened in chunks rather than immersing the reader in the story. GPT is slightly better at balancing the weight of the sentence, using passive language to make the narrator fade a bit more into the background ("As I sat...") and in return brings more focus to the woman's words. Overall, I think that when the sentence is translated as-is from its original form into English, the length itself is a bit of a problem. The detail that the woman is lying on her back feels like it was shoved into a sentence already packed with information. I don't know if this sounds more natural in Japanese or not (or if this is just a flaw in the original sentence itself) but if I were translating it, I might try and find a way to shift that detail around to the sentence before or after, even though this would be unfaithful to the way the text was written. I could just be nitpicking though.

In my second attempt, I used the sentence: 「行灯 も 蕪 村の 画も、畳も、違 棚も有って無いような、無くって有るように見えた。」

ChatGPT gave me: "The lantern, Buson’s painting, the tatami, even the staggered shelves — they all seemed to be there and not there, not there and yet somehow present."

Google Gemini said: "The andon lamp, the Buson painting, the tatami mat, and the chigai-dana shelf all seemed to both exist and not exist simultaneously."

Once again, GPT reigned supreme. I actually quite like the translation it gave, for its ability to portray the delicacy of the philosophical statement given at the end. I think it could still be cleaned up a bit (maybe "there but not there, non-existent and yet somehow present")? Gemini once again sounds robotic ... "[they] all seemed to both exist and not exist simultaneously" sounds overly simplistic and like the AI engine isn't aware that a philosophical lens is even being utilized. I also notice that GPT chose a different approach to translating the objects in the room, simplifying 行灯 to lantern when Gemini called it an "andon lamp," favoring a more literal and explanatory method. GPT even describes 違 棚 as "staggered shelves" while Gemini is overt in its naming of the "chigai-dana shelf." I suppose either way works, though I personally favor GPT's method a bit more - leaving in the literal Japanese name of the item before the English one feels a bit auditorily clunky to me even though it could be useful for the reader to paint a visual picture of specifically what style of lamp/shelf is being mentioned.

Overall, I sense from this exercise that AI machines can easily struggle with artistic nuance, as seen with the philosophical example in the second sentence. If I had chosen more dialogue-heavy passages, I think this issue would have been even more prominent. The wording of the sentences was often over-exact/literal, not caring to adjust its diction to sound more natural or attempt any sort of literary voice. My knowledge of Japanese is not good enough yet for me to be able to pick up on voice myself, but I have a feeling that this could be another pitfall of machine learning when it's used for translation.