Code for applying dynamic data selection as described in this paper
@inproceedings{vanderwees2017dynamic,
author = {van der Wees, Marlies and Bisazza, Arianna and Monz, Christof},
title = {Dynamic Data Selection for Neural Machine Translation},
booktitle = {Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing},
year = {2017},
pages = {1400--1410}
}
Before ranking the bitext files, four language models have to be created:
- a small in-domain LM in the source language
- a small in-domain LM in the target language
- a small general-domain LM in the source language
- a small general-domain LM in the target language
Note that when generating the general-domain LMs, the vocabulary is restricted to words that occur at least twice in the small in-domain corpus (following Axelrod et al.).
The LMs can be for example ngram LMs (e.g. created using SRILM) or LSTMs (I created these using functions from Tardis).
Next, compute for each of the above the cross-entropy of each sentence in the bitext with the in-domain and general-domain LMs. Make sure that the scores are written to files with one score per line (corresponding to the sentences in the bitext).
Bilingual cross-entropy difference (CED)/Modified Moore-Lewis, as presented by Axelrod et al., Domain Adaptation via Pseudo In-Domain Data Selection, 2011
$ python scripts/rank-bitext.py --bitext_files=data/bitext.src,data/bitext.trg --src_domain_loss=LM-I-SRC.blm --trg_domain_loss=LM-I-TRG.blm --src_general_loss=LM-G-SRC.blm --trg_general_loss=LM-G-TRG.blm
Requires:
- Four files with one cross-entropy score per line (see section above)
- Plain text bitext files with one sentence per line. At least two files are expected (e.g., bitext.src and bitext.trg). Additional files with for example meta-info can be added as well and will be ranked according to the same criteria.
Produces:
- Files with ranked bitext sentences, the most domain-relevant sentences on top
- File
ranked-bitext.weights: CED scores, one score per line, corresponding to sentence pairs in the bitext
Requires:
- ranked bitext files: train.src, train.trg, containing one sentence per line.
- specification of the DDS variant (sampling or gradual fine-tuning (gft))
- if applying sampling: a file with cross-entopy difference (CED) scores, one score per line corresponding to sentence pairs in the bitext.
Generates:
- training files for each epoch: train.src.1, train.trg.1, train.src.2, train.trg.2
In order to use these training files, modify your favorite NMT code such that it reads in new training files for each epoch, while using source and target vocabulary of the complete set of training data used in all epochs.
$ python scripts/dynamic-data-selection.py --bitext_src=data/train.src --bitext_trg=data/train.trg --dds_method=gft --alpha=1 --beta=0.8 --eta=1 --total_epochs=12
Default parameter values that were used in the paper:
$ --alpha=0.5 --beta=0.7 --eta=2 --total_epochs=16
$ python scripts/dynamic-data-selection.py --bitext_src=data/train.src --bitext_trg=data/train.trg --dds_method=sampling --ced_weights=data/weights.txt --alpha=1 --sampling_fraction=0.3 --total_epochs=12
Default parameter values that were used in the paper:
$ --alpha=0.5 --sampling_fraction=0.2 --total_epochs=16