Friday, September 4, 2026

Machine Translation

Large language models (LLMs) and other AI systems generally produce better translations when they are given more context. However, how well can they translate when only a short sentence or passage is provided? To explore this question, I will use the opening passage of Yasunari Kawabata’s novel Yukiguni (Snow Country) as a test case:

Japanese: 国境の長いトンネルを抜けると雪国であった。夜の底が白くなった。信号所に汽車が止まった。

English Translation by Edward Seidensticker: The train came out of the long tunnel into the snow country. The earth lay white under the night sky. The train pulled up at a signal stop.

Chatgpt 5.6 Sol Ultra input prompt: 

Translate the Japanese passage below into natural English. Base your translation only on the text provided, without consulting or reproducing an existing published translation. Preserve the meaning, tone, imagery, and relationships between the sentences. 

国境の長いトンネルを抜けると雪国であった。夜の底が白くなった。信号所に汽車が止まった。

Chatgpt 5.6 Sol Ultra output:

“Beyond the long tunnel at the border lay a land of snow. The depths of the night had turned white. The train stopped at a signal post.”

To see a difference, I also tested the prompt with a longer passage: 

Japanese: 国境の長いトンネルを抜けると雪国であった。夜の底が白くなった。信号所に汽車が止まった。

内側の座席から娘が立って来て、島村の前のガラス窓を落とした。雪の冷気が流れ込んだ。娘は窓いっぱいに乗り出して、遠くへ叫ぶように、

「駅長さん、駅長さん」

明かりをさげてゆっくり吹きを踏んできた男は、襟巻で鼻の上まで包み、耳に帽子の毛皮を垂れていた。

もうそんな寒さかと島村は外を眺めると鉄道の官舎らしいバラックが山裾に寒々と散らばっているだけで、雪の色はそこまで行かぬうちに闇に飲まれていた。

English Translation by Edward Seidensticker:

    The train came out of the long tunnel into the snow country. The earth lay white under the night sky. The train pulled up at a signal stop. 

    A girl who had been sitting on the other side of the car came over and opened the window in front of Shimamura. The snowy cold poured in. Leaning far out the window, the girl called to the station master as though he were a great distance away. 

    The station master walked slowly over the snow, a lantern in his hand. His face was buried to the nose in a muffler, and the flaps of his cap were turned down over his ears. 

    It's that cold, is it, thought Shimamura. Low, barracklike buildings that might have been railway dormitories were scattered here and there up the frozen slope of the mountain. The white of the snow fell away into the darkness some distance before it reached them.

Chatgpt 5.6 Sol Ultra input prompt: 

Translate the Japanese passage below into natural English. Base your translation only on the text provided, without consulting or reproducing an existing published translation. Preserve the meaning, tone, imagery, and relationships between the sentences. After translating it, briefly explain any words or expressions that are ambiguous without additional context.

国境の長いトンネルを抜けると雪国であった。夜の底が白くなった。信号所に汽車が止まった。

内側の座席から娘が立って来て、島村の前のガラス窓を落とした。雪の冷気が流れ込んだ。娘は窓いっぱいに乗り出して、遠くへ叫ぶように、

「駅長さん、駅長さん」

明かりをさげてゆっくり吹きを踏んできた男は、襟巻で鼻の上まで包み、耳に帽子の毛皮を垂れていた。

もうそんな寒さかと島村は外を眺めると鉄道の官舎らしいバラックが山裾に寒々と散らばっているだけで、雪の色はそこまで行かぬうちに闇に飲まれていた。

Chatgpt 5.6 Sol Ultra output:

Beyond the long tunnel at the border, everything was snow country. The depths of the night had turned white. The train stopped at a signal post. 

A young woman rose from a seat farther inside the carriage, came over, and lowered the glass window in front of Shimamura. The chill air off the snow streamed in. Filling the whole window as she leaned out, she called as though to someone far away:

“Stationmaster! Stationmaster!”

A man approached slowly with a light hanging from his hand, treading through the windblown snow. His scarf was wrapped above his nose, and the fur flaps of his cap hung over his ears. 

Was it already cold enough for all that? Shimamura looked outside. There was nothing but a bleak scattering of barrack-like buildings—apparently railway staff quarters—at the foot of the mountain, and the whiteness of the snow was swallowed by the darkness before it reached even that far.

-----------------------------------------------------------------------------------------------------------------------------
The results show that the longer passage had only a small effect on ChatGPT’s translation of the opening. In the shorter test, it translated the first sentence as 
“Beyond the long tunnel at the border lay a land of snow,” 
whereas the longer passage produced 
“Beyond the long tunnel at the border, everything was snow country.” 
The model appears to have understood the basic setting from the shorter passage alone, adding more context made very little difference. 

Terry

No comments:

Post a Comment

11/8 Readings

Hibbett's speech helped me realize how personal the art of translation is. Many translators have an very strong relationship with the or...