Scripts for training and evaluation of low-resource NMT

[Andar]

Dear all,

I need to train and evaluate an NMT model for a low-resource language with NEMATUS. Could you help me find the scripts for training and evaluation of experiments done in Revisiting Low-Resource Neural Machine Translation: A Case Study (https://arxiv.org/abs/1905.11901)?

Regards,

Andar

[Rico Sennrich]

Hi Andar,

You can start with the scripts in http://data.statmt.org/wmt17_systems/training/

You can then skip steps 2-3 (data download and preprocessing, and use the data from here instead: http://data.statmt.org/rsennrich/iwslt14_lowresource/), and start with step 4: just modify scripts/train.sh with the config options you want to test.

scripts/evaluate.sh uses multi-bleu-detok.perl . In the paper, we report results with https://github.com/mjpost/sacreBLEU , but both should give the same results.

best wishes,

Rico

--
You received this message because you are subscribed to the Google Groups "Nematus Support" group.
To unsubscribe from this group and stop receiving emails from it, send an email to nematus-suppo...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/nematus-support/7074fd0a-958b-45b6-9ea8-7d9b290ef03e%40googlegroups.com.

[Andar]

Hi Rico,

Thank you for your reply. Sadly, I couldn't replicate the experiment as Nematus requires Python 3.5.2 and Tensorflow 1.12. By this time, it is too difficult to install both even with anaconda. Is there any alternative?

Best,

Andar

On Wednesday, October 30, 2019 at 11:20:10 AM UTC+3, Rico Sennrich wrote:

Hi Andar,

You can start with the scripts in http://data.statmt.org/wmt17_systems/training/

You can then skip steps 2-3 (data download and preprocessing, and use the data from here instead: http://data.statmt.org/rsennrich/iwslt14_lowresource/), and start with step 4: just modify scripts/train.sh with the config options you want to test.

scripts/evaluate.sh uses multi-bleu-detok.perl . In the paper, we report results with https://github.com/mjpost/sacreBLEU , but both should give the same results.

best wishes,

Rico

On 30/10/2019 08:41, Andar wrote:

Dear all,

I need to train and evaluate an NMT model for a low-resource language with NEMATUS. Could you help me find the scripts for training and evaluation of experiments done in Revisiting Low-Resource Neural Machine Translation: A Case Study (https://arxiv.org/abs/1905.11901)?

Regards,

Andar

--
You received this message because you are subscribed to the Google Groups "Nematus Support" group.

To unsubscribe from this group and stop receiving emails from it, send an email to nematus...@googlegroups.com.

[Rico Sennrich]

Hi Andar,

apologies if the README was confusing. In fact, any newer Tensorflow version of the 1.X branch should be fine (just tested with TF 1.12, 1.13, 1.14 and 1.15), as well as newer Python versions.

best wishes,

Rico

To unsubscribe from this group and stop receiving emails from it, send an email to nematus-suppo...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/nematus-support/324bfd90-f20b-44cb-95bf-cba6decd8d84%40googlegroups.com.

[Andar]

Dear Rico,

Thanks a lot for your reply. I was running the attached bash scripts, train-lr.sh and validate-lr.sh. I still get the attached error message. I was using Tensorflow 1.14.0 and Python 3.6.9. Could you give me your comments?

Best,

Andar

To view this discussion on the web visit https://groups.google.com/d/msgid/nematus-support/324bfd90-f20b-44cb-95bf-cba6decd8d84%40googlegroups.com.

[Rico Sennrich]

Hello Andar,

I said you can skip preprocessing, but you'll actually need to execute the last step in the preprocessing script (building the network dictionary). Just run the following (either for both the source or target side, or, if you want to use tied embeddings, for the concatenation of both:

nematus/data_build_dictionary {corpus_name}

This will create files {corpus_name}.json, which you then pass to the "--dictionaries" argument when you call the Nematus training script.

best wishes,
Rico

To unsubscribe from this group and stop receiving emails from it, send an email to nematus-suppo...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/nematus-support/b773719c-f967-46d9-83db-593fe781b5c2%40googlegroups.com.

[Andar]

Dear Rico,

Thank you so much. It works perfectly. One more question. How can I set the BiDeep RNN architecture in the bash script? There are only two choices, rnn and transformer, for model_type.

Best,

Andar

To view this discussion on the web visit https://groups.google.com/d/msgid/nematus-support/b773719c-f967-46d9-83db-593fe781b5c2%40googlegroups.com.

[Rico Sennrich]

Hi Andar,

the BiDeep architecture is an RNN architecture, and the result of enabling two kinds of depth: stack depth and deep transition depth.

--rnn_enc_depth and --rnn_dec_depth control the number of layers (stack depth),
--rnn_enc_transition_depth, --rnn_dec_base_transition_depth and --rnn_dec_high_transition_depth control the number of GRU transitions per layer (deep transition depth).

best wishes,
Rico

To unsubscribe from this group and stop receiving emails from it, send an email to nematus-suppo...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/nematus-support/8aee6152-f199-46dd-bd29-a6ab069e8b9b%40googlegroups.com.