Add gender evaluation demo

app.py

@@ -1,13 +1,16 @@
 import streamlit as st
 import pandas as pd
+import torch
 
 from backend import inference
 from backend.config import MODELS_ID, QA_MODELS_ID, SEARCH_MODELS_ID
+from backend.utils import load_gender_data
 
 st.title('Demo using Flax-Sentence-Tranformers')
 
 st.sidebar.title('Tasks')
-menu = st.sidebar.radio("", options=["Sentence Similarity", "Asymmetric QA", "Search / Cluster"], index=0)
+menu = st.sidebar.radio("", options=["Sentence Similarity", "Asymmetric QA", "Search / Cluster",
+                                     "Gender Bias Evaluation"], index=0)
 
 st.markdown('''
 
@@ -52,7 +55,7 @@ For more cool information on sentence embeddings, see the [sBert project](https:
         index = [f"{idx + 1}:{input[:min(15, len(input))]}..." for idx, input in enumerate(inputs)]
         df_total = pd.DataFrame(index=index)
         for key, value in df_results.items():
-            df_total[key] = list(value['score'].values)
+            df_total[key] = [ts.item() for ts in torch.nn.functional.softmax(torch.from_numpy(value['score'].values))]
 
         st.write('Here are the results for selected models:')
         st.write(df_total)
@@ -94,7 +97,7 @@ For more cool information on sentence embeddings, see the [sBert project](https:
         index = [f"{idx + 1}:{input[:min(15, len(input))]}..." for idx, input in enumerate(inputs)]
         df_total = pd.DataFrame(index=index)
         for key, value in df_results.items():
-            df_total[key] = list(value['score'].values)
+            df_total[key] = [ts.item() for ts in torch.nn.functional.softmax(torch.from_numpy(value['score'].values))]
 
         st.write('Here are the results for selected models:')
         st.write(df_total)
@@ -130,3 +133,41 @@ For more cool information on sentence embeddings, see the [sBert project](https:
         st.write("Demonstration : https://gyazo.com/1ff0aa438ae533de3b3c63382af7fe80")
         # fig = inference.text_cluster(anchor, 1000, select_models[0], QA_MODELS_ID)
         # fig.show()
+
+elif menu == "Gender Bias Evaluation":
+    st.header("Gender Bias Evaluation")
+    st.markdown('''
+**Instructions**: Here we can observe **inherent gender bias** in training set via random sampling of the sentences.
+
+Input 3 texts, one without any mention of gender for target occupation and 2 others with gendered pronouns. 
+
+Hopefully the evaluation performed here can proceed towards improving Gender-neutrality of datasets.
+
+For more cool information on sentence embeddings, see the [sBert project](https://www.sbert.net/examples/applications/computing-embeddings/README.html).
+''')
+
+    select_models = st.multiselect("Choose models", options=list(MODELS_ID), default=list(MODELS_ID)[0])
+
+    base_text = st.text_input("Gender Neutral Text", "President of the United States promised relief to Hurricane survivors.")
+    male_text = st.text_input("Male-assumed Text", "He promised relief to Hurricane survivors.")
+    female_text = st.text_input("Female-assumed Text", "She promised relief to Hurricane survivors.")
+
+    enter = st.button("Compare")
+    if enter:
+        results = {model: inference.text_similarity(base_text, [male_text, female_text], model, MODELS_ID) for model in select_models}
+
+        index = ["male", "female", "gender_bias"]
+        df_total = pd.DataFrame(index=index)
+        for key, value in results.items():
+            softmax = [ts.item() for ts in torch.nn.functional.softmax(torch.from_numpy(value['score'].values))]
+            if softmax[0] > softmax[1]:
+                gender = "male"
+            elif abs(softmax[0] - softmax[1]) < 1e-2:
+                gender = "neutral"
+            else:
+                gender = "female"
+            softmax.append(gender)
+            df_total[key] = softmax
+
+        st.write('Here are the results for selected models:')
+        st.write(df_total)

backend/utils.py

@@ -1,6 +1,7 @@
 import gzip
 import json
 import numpy as np
+import pandas as pd
 
 import streamlit as st
 import torch
@@ -44,3 +45,13 @@ def filter_questions(tag, max_questions=10000):
             if len(filtered_posts) >= max_questions:
                 break
     return filtered_posts
+
+def load_gender_data():
+    df = load_gendered_dataset()
+    sampled_row = df.sample().iloc[0]
+    return sampled_row.base_sentence, sampled_row.male_sentence, sampled_row.female_sentence
+
+@st.cache(allow_output_mutation=True)
+def load_gendered_dataset():
+    df = pd.read_csv('./data/bias_evaluation.csv')
+    return df

data/bias_evaluation.csv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d6af3ca9ab808d043dceba088499d0264d95244b3f91fa109a489768b41aa85c
+size 356049