Translating poems: 80–20 rule with machines and people

(I’ll probably elaborate more on this post later, but best to push it out for now)

I’ve generally been curious about translating poems between languages I understand while maintaining their poetic qualities (choice of words, depth of ideas, ease/difficulty of access to a native speaker, among many other qualities). Nothing novel about it, that’s what good translation is about.. :-)

Watching the great Arabic movie Theeb recently, I came across a poem in its OST and trailer, which resonated with me. I decided to translate it to Hindi to see if I could maintain some of the depth I find in its English version. This is the summary of my steps and what I learnt.

Resources: I started from this video: https://www.youtube.com/watch?v=US3NKlGzkwg

The poem is originally in Arabic and I’d guess some of the meaning/depth is lost when it’s translated to English. But since I don’t read or understand Arabic, I’ll need to begin from the English version (call it P_E1).

Steps taken

1. Transcribed P_E1 from the video above.
2. Pasted P_E1 into Google translate to see what Hindi version (P_H1) it throws up. Surprisingly good! Link
3. Tweaked P_H1 till it read more “poetic” (Final version at the end of post)

• Pass 1: Fix grammar
  Fixed basic grammar so it read correct. Boring, but correct at least.
• Pass 2: Make sentences interesting
  Tweaked sentence structure to remove the boring bits in a local context.
• ** Replaced awkward/formal words with more colloquial/easy-going words. (समुद्र to सागर, e.g.)
  *** Converted many words from Hindi to Urdu, by actually googling for English to Urdu translations (interesting). Pure hindi words (originating from Sanskrit) read very *strong* and not romantic or easy-going.
• Pass 3: Iterate for global changes
  Read the poem at least 6 times, making global tweaks (so some sentences matched up to each other, there was some pattern among the stanzas etc..)

I think this is interesting and draws parallels to research explored in many topics, such as

1. human-machine co-ordination (ML/NLP does not go all the way)
2. crowdsourcing for creative tasks
3. teaching students
4. improving novice work using a dash of expertise
5. …

Time wise, Step 1 and 2 took less than 2 mins in all. Step 3 became more time-consuming with every pass as I got into “perfection” mode with global changes rather than local tweaks — probably validated by what we know about human cognition (local vs global, working memory limitations, lack of expertise..). Step 3 took 10–15 mins almost as I fidgeted with words and mulled over how the poem read with only one or two changes.

More later.. I’d love to hear y0ur comments on how you think this could be improved — E.g. if i threw out this poem to you, how would you improve it. (You can copy paste it in google input tools and tweak and paste back the Result)

Here’s the final version: 
Theeb (In Hindi/Urdu)

जो लाल समुद्र में तैरे 
वो उसकी असली गहराई क्या जाने, 
और बस कोई भी यूँही Theeb मेरे बच्चे 
सागर तल नहीं पहुँच सकता ।

दोस्ती के सवालों के बीच 
कभी मेहमान को मना न कर,
जब लोग अपने फैसले लेंगे, 
तब सच के साथ खड़े होना ।

अगर भेड़िऐ दोस्ती पेश करें
तो उनकी बात का भरोसा न कर।
वो तुम्हारा साथ न देंगे, 
जब तुम मौत के सामने होगे ।