The issue was that Moses was executed after pre_rules when use_moses = True, But when data set was pre tokenized with Moses (use_moses=False) the pre_rules were executed after.
So our wikipedia had the following processing:
- raw text
- Moses
- pre_rules
- split(' ') # fastai BaseTokenizer
- post_rules
- sentence piece
While mldoc had the following tokenziation
- raw text
- pre_rules
- Moses
- post_rules
- sentence piece
After fix I've retrained the classfiers (without finetuning) and I haven't notice huge changes in the performance. 4 languages received slight improvment 4 got a slight decrease in performance.
f'xx{token_name}' are kept as one token by Moses tokenizer, which is sometimes required if you want to have moses in tokenizers pipeline, and we use that for imdb.
- train on whole articles when tokenizer is sentencepice
- add moses tokenizer to get the same tokens on imdb as on wt103
- apply pre / post processing rules to moses tokenzier so that sentencepiece works on correctly preprocessed text (lowercased with html removed)
- add alpha implementation on xnli (no tests)
- remove vocab adaptation when swiching from wiki to imdb. As otherwise we get 50% of missing words and 93% accuracy on imdb
Piotr Czapla