cartesinus/iva_mt_wslot-m2m100_418M-en-pl

iva_mt_wslot-m2m100_418M-0.1.0 en-pl

This model is a fine-tuned version of facebook/m2m100_418M on the iva_mt_wslot dataset. It achieves the following results:

1. On the test set (iva_mt):

• BLEU (plain text): 39.1560
• BLEU (with slots): 63.8767
• F1 score: (in preparation)

For reference BLEU for baseline m2m100-418M (plain text) was 21.9468. Second result (63.8767) is when tags are treated as "normal" words in sentence. Therefore that result might be a bit misleading. Please refer to plain text results if you are not sure how to interpret them.

2. WMT20 (en2pl):

• BLEU (lowercased, tags removed): 15.0863
• BLEU for baseline m2m100-418M (plain text): 20.2750

For reference WMT20 submission systems in en-pl direction had between 25 and 30 BLEU

3. BLEU on the evaluation set (same as in below table 'Training results'): 61.6249

4. On the training set (to see how it adjusted to train):

• BLEU (plain text): 70.5597
• BLEU (with slots): 93.8200

BLEU was measured with sacreBLEU library.

Model description, intended uses & limitations

Model is biased towards virtual assistant (IVA) sentences in prediction/translation. These sentences are short, imperatives with a lot of name entities (slots) and particular vocabulary (for example settings name). It can be observed in above results where WMT results are very low while in-domain test is very high.

This model will most probably force IVA translations on your text. As long as sentences that you are translating are more or less similar to massive and leyzer domains it will be ok. If you will translate out-of-domain sentenences (such as for example News, Medical) that are not very similar then results will drop significantly up to the point where baseline m2m100-418M will be better than this model.

This model will generate tags in output even if there is not tag in input sentence. Frequency of this depends on input text origin. When testing IVA utterances this occurs between 3 and 5%. When WMT20 was translated it happened in 40% cases (input text was from News domain). This is not very severe problem and it can be fixed easily in post-processing (simple sed 's/<[a-z]>//g' should be enough in most cases).

Translations with slot annotation very often differ from same sentences when slots are removed. This is quite frequent and it happens between 30 and 50% of translated utterances. For example there will be a difference between "is it raining in barcelona" and "is it raining in <a>barcelona<a>". In second case model will more likely localize name of city to some Polish name (here Lublin, because such city was given in Massive train set). This might be useful if you want to generate more variants.

One last thing that needs to be mentioned is that BLEU is not particulary good metric to evaluate IVA sentences due to their length and it should be evalued with other metrices (e.g. GLEU).

How to use

First please make sure to install pip install transformers. First download model:

from transformers import M2M100ForConditionalGeneration, M2M100Tokenizer
import torch

def translate(input_text, lang):
    input_ids = tokenizer(input_text, return_tensors="pt")
    generated_tokens = model.generate(**input_ids, forced_bos_token_id=tokenizer.get_lang_id(lang))
    return tokenizer.batch_decode(generated_tokens, skip_special_tokens=True)

model_name = "cartesinus/iva_mt_wslot-m2m100_418M-0.1.0"
tokenizer = M2M100Tokenizer.from_pretrained(model_name, src_lang="en", tgt_lang="pl")
model = M2M100ForConditionalGeneration.from_pretrained(model_name)

Then you can translate either plan text like this:

print(translate("set the temperature on my thermostat", "pl"))

or you can translate with slot annotations that will be restored in tgt language:

print(translate("wake me up at <a>nine am<a> on <b>friday<b>", "pl")) #translation: obudź mnie o <a>piątej rano<a> <b>w tym tygodniu<b>

Limitations of translation with slot transfer:

1. Annotated words must be placed between semi-xml tags like this "this is <a>example<a>"
2. There is no closing tag for example "<\a>" in above example - this is done on purpose to ommit problems with backslash escape
3. If sentence consists of more than one slot then simply use next alphabet letter. For example "this is <a>example<a> with more than <b>one<b> slot"
4. Please do not add space before first or last annotated word because this particular model was trained this way and it most probably will lower it's results

Training and evaluation data

Dataset Composition (en-pl)

Corpus	Train	Dev	Test
Massive 1.1	11514	2033	2974
Leyzer 0.2.0	3974	701	1380
OpenSubtitles from OPUS	2329	411	500
KDE from OPUS	1154	241	241
CCMatrix from Opus	1096	232	237
Ubuntu from OPUS	281	60	59
Gnome from OPUS	14	3	3
total	20362	3681	5394

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 2e-05
• train_batch_size: 4
• eval_batch_size: 4
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• num_epochs: 10
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	BLEU	Gen Len
0.0178	1.0	5091	0.0171	57.4439	21.1396
0.013	2.0	10182	0.0159	58.886	21.2285
0.0091	3.0	15273	0.0157	60.159	21.1222
0.0073	4.0	20364	0.0159	60.5893	21.1212
0.0054	5.0	25455	0.0161	60.6484	21.0679
0.004	6.0	30546	0.0166	61.5283	21.0875
0.0031	7.0	35637	0.0169	61.0439	21.1562
0.0024	8.0	40728	0.0172	61.9427	21.2203
0.0018	9.0	45819	0.0175	61.7325	21.1478
0.0014	10.0	50910	0.0176	61.6249	21.157

Framework versions

• Transformers 4.26.1
• Pytorch 1.13.1+cu116
• Datasets 2.10.1
• Tokenizers 0.13.2

Citation

If you use this model, please cite the following:

@article{Sowanski2023SlotLI,
  title={Slot Lost in Translation? Not Anymore: A Machine Translation Model for Virtual Assistants with Type-Independent Slot Transfer},
  author={Marcin Sowanski and Artur Janicki},
  journal={2023 30th International Conference on Systems, Signals and Image Processing (IWSSIP)},
  year={2023},
  pages={1-5}
}