We are trying to create Portuguese-English translation model. It’s based on open source materials ~6M.
I’ve created 500K/400K PT/EN vocabulary. And model was trained with 2x1000 neurons for encoder/decoder.
Residual TRUE, encoder type GNMT. In the end of 11 epochs perplexity was ~4.3.
And model works fine in many cases BUT I’ve got empty translations for some strings. Is that looks like a bug, or I miss something?
Thanks !
[04/04/18 13:34:46 INFO] SENT 16: b ) a projeção de investimentos para a implantação e manutenção do projeto ;
[04/04/18 13:34:46 INFO] PRED 16: b )
[04/04/18 13:34:46 INFO] PRED SCORE: -4.04
[04/04/18 13:34:46 INFO] SENT 17: c ) as estimativas de geração de empregos diretos e indiretos e de recolhimento de tributos diretos e indiretos ;
[04/04/18 13:34:46 INFO] PRED 17: c )
[04/04/18 13:34:46 INFO] PRED SCORE: -4.77
My guess is that your training data have few or no examples containing single closing parentheses without matching opening parentheses. So when the model sees that at inference, it has some trouble to deal with it.
We have translated strings with similar cases, that’s strange:
[04/05/18 10:33:39 INFO] SENT 15: a ) caracterização do projeto , com a indicação da ( s ) cultura ( s ) a ser ( em ) implantada ( s ) , densidade por hectare , objetivo do plantio e as intervenções eventualmente necessárias à sua implantação e operação ;
[04/05/18 10:33:39 INFO] PRED 15: the profiling of the project , with a ( s ) culture ( s ) being ( s ) being ( in ) ( s ) , density per hectare , the aim of planting and the interventions that may be necessary for its deployment and operation ;
[04/05/18 10:33:39 INFO] PRED SCORE: -25.40
[04/05/18 10:33:39 INFO]
[04/05/18 10:33:39 INFO] SENT 16: b ) a projeção de investimentos para a implantação e manutenção do projeto ;
[04/05/18 10:33:39 INFO] PRED 16: b )
[04/05/18 10:33:39 INFO] PRED SCORE: -4.04
[04/05/18 10:33:39 INFO]
[04/05/18 10:33:39 INFO] SENT 17: c ) as estimativas de geração de empregos diretos e indiretos e de recolhimento de tributos diretos e indiretos ;
[04/05/18 10:33:39 INFO] PRED 17: c )
[04/05/18 10:33:39 INFO] PRED SCORE: -4.77
[04/05/18 10:33:39 INFO]
[04/05/18 10:33:39 INFO] SENT 18: d ) o aproveitamento racional e adequado da área rural adquirida , observada a legislação aplicável no que tange às áreas de preservação permanente e reserva legal ;
[04/05/18 10:33:39 INFO] PRED 18: ( d ) The rational and appropriate utilisation of the rural area acquired , observed the applicable legislation in the areas of permanent preservation and legal reservation ;
[04/05/18 10:33:39 INFO] PRED SCORE: -14.28
[04/05/18 10:33:39 INFO]
[04/05/18 10:33:39 INFO] SENT 19: e ) os impactos econômicos e sociais estimados do projeto para o desenvolvimento da região na qual será implantado ; e
[04/05/18 10:33:39 INFO] PRED 19: and the estimated economic and social impacts of the project for the development of the region in which it will be deployed ; and
[04/05/18 10:33:39 INFO] PRED SCORE: -6.09