Machine Translation Explanations

This notebook demonstrates model explanations for a text to text scenario using a pretrained transformer model for machine translation. In this demo, we showcase explanations on two different models: English to Spanish (https://huggingface.co/Helsinki-NLP/opus-mt-en-es), and English to French (https://huggingface.co/Helsinki-NLP/opus-mt-en-fr).

[1]:

from transformers import AutoModelForSeq2SeqLM, AutoTokenizer

import shap

English to Spanish model

[2]:

# load the model and tokenizer
tokenizer = AutoTokenizer.from_pretrained("Helsinki-NLP/opus-mt-en-es")
model = AutoModelForSeq2SeqLM.from_pretrained("Helsinki-NLP/opus-mt-en-es").cuda()

# define the input sentences we want to translate
data = [
    "Transformers have rapidly become the model of choice for NLP problems, replacing older recurrent neural network models"
]

Explain the model’s predictions

[3]:

# we build an explainer by passing the model we want to explain and
# the tokenizer we want to use to break up the input strings
explainer = shap.Explainer(model, tokenizer)

# explainers are callable, just like models
shap_values = explainer(data, fixed_context=1)

floor_divide is deprecated, and will be removed in a future version of pytorch. It currently rounds toward 0 (like the 'trunc' function NOT 'floor'). This results in incorrect rounding for negative values.
To keep the current behavior, use torch.div(a, b, rounding_mode='trunc'), or for actual floor division, use torch.div(a, b, rounding_mode='floor'). (Triggered internally at  /pytorch/aten/src/ATen/native/BinaryOps.cpp:467.)

Visualize shap explanations

[4]:

shap.plots.text(shap_values)

[0]
outputs
Los
transformador
es
se
han
convertido
rápidamente
en
el
modelo
de
elección
para
problemas
N
LP
,
reemplaza
ndo
modelos
de
red
neuro
nal
recurrente
s
más
antiguos


-0-1010-7.50297-7.50297base value0.7807650.780765fLos(inputs)5.114 ers 1.965 ▁Transform 1.903 ▁have 0.325 ▁of 0.186 ▁become 0.172 ▁neural 0.105 ▁network 0.101 ▁the 0.081 ▁for 0.053 ▁replacing 0.05 ▁recurrent 0.025 ▁older 0.021 LP -0.505 ▁rapidly -0.417 , -0.247 ▁problems -0.225 ▁model -0.114 ▁choice -0.114 ▁models -0.1 -0.096 ▁N
inputs
1.965
▁Transform
5.114
ers
1.903
▁have
-0.505
▁rapidly
0.186
▁become
0.101
▁the
-0.225
▁model
0.325
▁of
-0.114
▁choice
0.081
▁for
-0.096
▁N
0.021
LP
-0.247
▁problems
-0.417
,
0.053
▁replacing
0.025
▁older
0.05
▁recurrent
0.172
▁neural
0.105
▁network
-0.114
▁models
-0.1