First initial commit

README.md

@@ -1,13 +1,12 @@
 ---
-title: MT Bias Demo
-emoji: 💩
-colorFrom: green
-colorTo: red
+title: Bias in MT
+emoji: 🌍
+colorFrom: yellow
+colorTo: indigo
 sdk: gradio
-sdk_version: 3.16.0
+sdk_version: 3.3
 app_file: app.py
 pinned: false
-license: mit
 ---
 
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+A demo showing how gender bias could manifest in MT models when translating from Hungarian to English.

app.py

@@ -0,0 +1,148 @@
+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()

contrastive_pair.md

@@ -0,0 +1 @@
+This example is very similar to the **Simple translation** example, but now we ask how the model's behaviour would change if we change the translation of “ő” from “he” to “she”? The `probability` row at the bottom shows the difference in the probability between both versions of the translation.

description.md

@@ -0,0 +1,9 @@
+# Gender Bias in MT: Hungarian to English
+
+The Hungarian language has no grammatical gender and words like “he” and “she” are both translated as “ő”.
+This makes it an interesting language to study gender bias in machine translation (MT) models, when translating to another language that does distinguish between “he” and “she”.
+In this demo, we will test the OPUS-MT models (Tiedemann & Thottingal, 2020) from the *Language Technology Research Group at the University of Helsinki* ([Helsinki-NLP](https://github.com/Helsinki-NLP)).
+
+For each translation, we also use the [Inseq library](https://github.com/inseq-team/inseq) to compute the feature attributions with integrated gradients: How important is each token in the source (Hungarian) for the translation of the target tokens (English)?
+
+⚠️ Please note that this demo is just an illustration of how gender bias could manifest in MT models, but an actual assessment of its bias requires a more rigourous experiment.

notice.md

@@ -0,0 +1,8 @@
+The idea for testing the gender bias in translations from Hungarian to English comes from Farkas and Németh (2022).
+
+### References:
+[Inseq: Intepretability for Sequence Generation Models 🔍](https://github.com/inseq-team/inseq). GitHub.
+
+Tiedemann, J., & Thottingal, S. (2020). [OPUS-MT — Building open translation services for the World](https://helda.helsinki.fi/handle/10138/327852). Proceedings of the 22nd Annual Conferenec of the European Association for Machine Translation (EAMT).
+
+Farkas, A., & Németh, R. (2022). [How to measure gender bias in machine translation: Real-world oriented machine translators, multiple reference points](https://www.sciencedirect.com/science/article/pii/S2590291121001352). Social Sciences & Humanities Open, 5(1), 100239.

simple_translation.md

@@ -0,0 +1,5 @@
+Select an occupation (or the word “woman”/“man”) from the dropdown menu and press `Translate & Attribute` to translate a sentence like “He/She is a nurse” from Hungarian:
+
+> Ő nővér.
+
+Which pronouns (“she”/“he”) do the MT models go for? Does it change depending on the occupation term you choose? And can we find a difference between the target languages (you can change it in the other dropdown menu on the right)?