Hugging Face's logo

Spaces:
oskarvanderwal
/
MT-bias-demo
Running

App Files Community
MT-bias-demo
File size: 4,836 Bytes 
97a1414
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
 
import gradio
import inseq
from inseq.data.aggregator import AggregatorPipeline, SubwordAggregator, SequenceAttributionAggregator, PairAggregator
import torch

if torch.cuda.is_available():
    DEVICE = "cuda"
else:
    DEVICE = "cpu"

def swap_pronoun(sentence):
    if "He" in sentence:
        return sentence.replace("He", "She")
    elif "She" in sentence:
        return sentence.replace("She", "He")
    else:
        return sentence

def run_counterfactual(occupation):
    occupation = occupation.split(" (")[0]

    model_name = f"Helsinki-NLP/opus-mt-hu-en"

    # "egy" means something like "a", but is used less frequently than in English.
    #source = f"Ő egy {occupation}."
    source = f"Ő {occupation}."
    
    model = inseq.load_model(model_name, "integrated_gradients")
    model.device = DEVICE
    target = model.generate(source)[0]
    #target_modified = swap_pronoun(target)

    out = model.attribute(
    [
        source,
        source,
    ],
    [
        #target,
        #target_modified,
        target.replace("She", "He"),
        target.replace("He", "She"),
    ],
    n_steps=150,
    return_convergence_delta=False,
    attribute_target=False,
    step_scores=["probability"],
    internal_batch_size=100,
    include_eos_baseline=False,
    device=DEVICE,
)
    #out = model.attribute(source, attribute_target=False, n_steps=150, device=DEVICE, return_convergence_delta=False, step_scores=["probability"])

    squeezesum = AggregatorPipeline([SubwordAggregator, SequenceAttributionAggregator])
    masculine = out.sequence_attributions[0].aggregate(aggregator=squeezesum)
    feminine = out.sequence_attributions[1].aggregate(aggregator=squeezesum)

    return masculine.show(aggregator=PairAggregator, paired_attr=feminine, return_html=True, display=True)
    #return out.show(return_html=True, display=True)

def run_simple(occupation, lang, aggregate):
    occupation = occupation.split(" (")[0]

    model_name = f"Helsinki-NLP/opus-mt-hu-{lang}"

    # "egy" means something like "a", but is used less frequently than in English.
    #source = f"Ő egy {occupation}."
    source = f"Ő {occupation}."
    
    model = inseq.load_model(model_name, "integrated_gradients")
    out = model.attribute([source], attribute_target=True, n_steps=150, device=DEVICE, return_convergence_delta=False)
    
    if aggregate:
        squeezesum = AggregatorPipeline([SubwordAggregator, SequenceAttributionAggregator])
        return out.show(return_html=True, display=True, aggregator=squeezesum)
    else:
        return out.show(return_html=True, display=True)

with open("description.md") as fh:
    desc = fh.read()

with open("simple_translation.md") as fh:
    simple_translation = fh.read()

with open("contrastive_pair.md") as fh:
    contrastive_pair = fh.read()

with open("notice.md") as fh:
    notice = fh.read()

OCCUPATIONS = [
    "nő (woman)",    
    "férfi (man)",
    "nővér (nurse)",
    "tudós (scientist)",
    "mérnök (engineer)",
    "pék (baker)",
    "tanár (teacher)",
    "esküvőszervező (wedding organizer)",
    "vezérigazgató (CEO)",
]

LANGS = [
    "en",
    "fr",
    "de",
]

with gradio.Blocks(title="Gender Bias in MT: Hungarian to English") as iface:
    gradio.Markdown(desc)

    print(simple_translation)
    with gradio.Accordion("Simple translation", open=True):
        gradio.Markdown(simple_translation)

    with gradio.Accordion("Contrastive pair", open=False):
        gradio.Markdown(contrastive_pair)

    gradio.Markdown("**Does the model seem to rely on gender stereotypes in its translations?**")

    with gradio.Tab("Simple translation"):
        with gradio.Row(equal_height=True):
            with gradio.Column(scale=4):
                occupation_sel = gradio.Dropdown(label="Occupation", choices=OCCUPATIONS, value=OCCUPATIONS[0])
            with gradio.Column(scale=4):
                target_lang = gradio.Dropdown(label="Target Language", choices=LANGS, value=LANGS[0])
        aggregate_subwords = gradio.Radio(
            ["yes", "no"], label="Aggregate subwords?", value="yes"
        )
        but = gradio.Button("Translate & Attribute")
        out = gradio.HTML()
        args = [occupation_sel, target_lang, aggregate_subwords]
        but.click(run_simple, inputs=args, outputs=out)

    with gradio.Tab("Contrastive pair"):
        with gradio.Row(equal_height=True):
            with gradio.Column(scale=4):
                occupation_sel = gradio.Dropdown(label="Occupation", choices=OCCUPATIONS, value=OCCUPATIONS[0])
        but = gradio.Button("Translate & Attribute")
        out = gradio.HTML()
        args = [occupation_sel]
        but.click(run_counterfactual, inputs=args, outputs=out)

    with gradio.Accordion("Notes & References", open=False):
        gradio.Markdown(notice)


iface.launch()