Spaces:
Sleeping
Sleeping
Create app.py
Browse files
app.py
ADDED
@@ -0,0 +1,164 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import streamlit as st
|
2 |
+
import pandas as pd
|
3 |
+
from sklearn.metrics import (
|
4 |
+
accuracy_score,
|
5 |
+
precision_score,
|
6 |
+
recall_score,
|
7 |
+
f1_score)
|
8 |
+
from imblearn.metrics import specificity_score
|
9 |
+
import difflib as dl
|
10 |
+
import os
|
11 |
+
|
12 |
+
|
13 |
+
# Title and description
|
14 |
+
st.title("Robustness and Sensitivity of BERT Models Predicting Alzheimer's Disease from Text")
|
15 |
+
st.markdown("Supplemantary material accompanying the following paper: Jekaterina Novikova (2021).[Robustness and Sensitivity of BERT Models Predicting Alzheimer's Disease from Text](https://arxiv.org/abs/2109.11888). \
|
16 |
+
*In: The 7th Workshop on Noisy User-generated Text at EMNLP*, 2021.", unsafe_allow_html=True)
|
17 |
+
st.image('img/poster2.png')
|
18 |
+
st.write("[Link](https://arxiv.org/abs/2109.11888) to the high-res version of the poster.")
|
19 |
+
|
20 |
+
# Loading data
|
21 |
+
my_data = "data/df_test_all.csv"
|
22 |
+
@st.cache(persist = True)
|
23 |
+
def load_data(dataset):
|
24 |
+
df = pd.read_csv(os.path.join(dataset))
|
25 |
+
return df
|
26 |
+
|
27 |
+
df = load_data(my_data)
|
28 |
+
|
29 |
+
# Sidebar to select type and level of perturbation selection menu
|
30 |
+
st.sidebar.title("Selection Menu")
|
31 |
+
st.sidebar.markdown("Please select the type and the level of text perturbation below. <hr>", unsafe_allow_html=True)
|
32 |
+
|
33 |
+
type = st.sidebar.selectbox('Type of perturbations', ["Original / No perturbations", "Delete filled pauses", "Delete info units", "Back-translation", "Substitute with WordNet synonyms"])
|
34 |
+
level = None
|
35 |
+
iu_type = None
|
36 |
+
|
37 |
+
if type in ["Substitute with word2vec", "Substitute with WordNet synonyms"]:
|
38 |
+
level = st.sidebar.slider('Level of perturbations:', min_value = 0.1, max_value = 0.90, step = 0.10)
|
39 |
+
elif type == "Delete info units":
|
40 |
+
iu_type = st.sidebar.radio('Type of info units:', ["Action only", "Location only", "Object only", "Subject only"])
|
41 |
+
|
42 |
+
|
43 |
+
# select column names based on subtype of perturbations:
|
44 |
+
def select_pred_column(type, level = None, iu_type = None):
|
45 |
+
if type == "Original / No perturbations":
|
46 |
+
prediction = "pred_original"
|
47 |
+
elif type == "Delete filled pauses":
|
48 |
+
prediction = "pred_no_filled_pause"
|
49 |
+
elif type == "Delete info units":
|
50 |
+
if iu_type == "Action only":
|
51 |
+
prediction = "pred_no_iu_action"
|
52 |
+
elif iu_type == "Location only":
|
53 |
+
prediction = "pred_no_iu_loc"
|
54 |
+
elif iu_type == "Object only":
|
55 |
+
prediction = "pred_no_iu_obj"
|
56 |
+
elif iu_type == "Subject only":
|
57 |
+
prediction = "pred_no_iu_subj"
|
58 |
+
elif type == "Back-translation":
|
59 |
+
prediction = "pred_back_transl"
|
60 |
+
elif type == "Substitute with word2vec":
|
61 |
+
lvl_str = str(level * 100)[:2]
|
62 |
+
prediction = "pred_w2v_"+lvl_str
|
63 |
+
elif type == "Substitute with WordNet synonyms":
|
64 |
+
lvl_str = str(level * 100)[:2]
|
65 |
+
prediction = "pred_wnet_"+lvl_str
|
66 |
+
return prediction
|
67 |
+
|
68 |
+
def select_aug_column(type, level = None, iu_type = None):
|
69 |
+
if type == "Original / No perturbations":
|
70 |
+
augmentation = "utterances"
|
71 |
+
elif type == "Delete filled pauses":
|
72 |
+
augmentation = "aug_no_filled_pause"
|
73 |
+
elif type == "Delete info units":
|
74 |
+
if iu_type == "Action only":
|
75 |
+
augmentation = "aug_no_iu_action"
|
76 |
+
elif iu_type == "Location only":
|
77 |
+
augmentation = "aug_no_iu_loc"
|
78 |
+
elif iu_type == "Object only":
|
79 |
+
augmentation = "aug_no_iu_obj"
|
80 |
+
elif iu_type == "Subject only":
|
81 |
+
augmentation = "aug_no_iu_subj"
|
82 |
+
elif type == "Back-translation":
|
83 |
+
augmentation = "aug_back_transl"
|
84 |
+
elif type == "Substitute with word2vec":
|
85 |
+
lvl_str = str(level * 100)[:2]
|
86 |
+
augmentation = "aug_w2v_"+lvl_str
|
87 |
+
elif type == "Substitute with WordNet synonyms":
|
88 |
+
lvl_str = str(level * 100)[:2]
|
89 |
+
augmentation = "aug_wnet_"+lvl_str
|
90 |
+
return augmentation
|
91 |
+
|
92 |
+
#part I
|
93 |
+
st.header("1. Classification Performance")
|
94 |
+
|
95 |
+
st.write("The performance of the fine-tuned BERT model tested on the samples of text with applied perturbations, as selected in the Selection Menu.")
|
96 |
+
|
97 |
+
if st.button("Calculate performance"):
|
98 |
+
acc = accuracy_score(df.label.values, df[select_pred_column(type, level, iu_type)].values)
|
99 |
+
f1 = f1_score(df.label.values, df[select_pred_column(type, level, iu_type)].values)
|
100 |
+
prec = precision_score(df.label.values, df[select_pred_column(type, level, iu_type)].values)
|
101 |
+
rec = recall_score(df.label.values, df[select_pred_column(type, level, iu_type)].values)
|
102 |
+
spec = specificity_score(df.label.values, df[select_pred_column(type, level, iu_type)].values)
|
103 |
+
|
104 |
+
df_perf = pd.DataFrame([acc, f1, prec, rec, spec])
|
105 |
+
df_perf.index = ["Accuracy", "F1-score", "Precision", "Recall/Sensitivity", "Specificity"]
|
106 |
+
df_perf.columns = ["Performance"]
|
107 |
+
st.table( df_perf.T)
|
108 |
+
|
109 |
+
#part II
|
110 |
+
st.header("2. Examples of Text Perturbations")
|
111 |
+
|
112 |
+
|
113 |
+
def text_to_code(text):
|
114 |
+
if text == "Healthy Control (label 0)":
|
115 |
+
code = [0]
|
116 |
+
elif text == "Alzheimer's Disease (label 1)":
|
117 |
+
code = [1]
|
118 |
+
else:
|
119 |
+
code = [0,1]
|
120 |
+
return code
|
121 |
+
|
122 |
+
dx = st.radio('Real disease:', ["Alzheimer's Disease (label 1)", "Healthy Control (label 0)", "both"])
|
123 |
+
pred1 = st.radio('Original prediction (before text perturbation):', ["Alzheimer's Disease (label 1)", "Healthy Control (label 0)", "Don't care"])
|
124 |
+
pred2 = st.radio('Prediction after text perturbation:', ["Alzheimer's Disease (label 1)", "Healthy Control (label 0)", "Don't care"])
|
125 |
+
|
126 |
+
subject_ids = df[(df["label"].isin(text_to_code(dx))) & \
|
127 |
+
(df["pred_original"].isin(text_to_code(pred1))) &\
|
128 |
+
(df[select_pred_column(type, level, iu_type)].isin(text_to_code(pred2)))]["subject_id"]
|
129 |
+
|
130 |
+
st.write('There are', subject_ids.shape[0], 'text sample(s) that correspond to such a selection.')
|
131 |
+
|
132 |
+
if subject_ids.shape[0] > 0:
|
133 |
+
subj_choice = st.selectbox("Select a text sample:", subject_ids)
|
134 |
+
|
135 |
+
df_select = df[df.subject_id == subj_choice][["subject_id", "sex", "age", "label", "pred_original", select_pred_column(type, level, iu_type)]]
|
136 |
+
df_select.age = df_select.age.astype(int)
|
137 |
+
df_select.columns = ["SubjectID", "Sex", "Age", "Real disease label", "Original prediction", "Prediction after perturbation"]
|
138 |
+
st.table(df_select)
|
139 |
+
|
140 |
+
text_orig = df[df.subject_id == subj_choice]["utterances"].values[0]
|
141 |
+
text_aug = df[df.subject_id == subj_choice][select_aug_column(type, level, iu_type)].values[0]
|
142 |
+
|
143 |
+
words_aug = set(text_aug.replace("'"," ' ").split())
|
144 |
+
words_orig = set(text_orig.replace("'"," ' ").split())
|
145 |
+
s1 = text_orig.replace("'"," ' ").split()
|
146 |
+
s2 = text_aug.replace("'"," ' ").split()
|
147 |
+
|
148 |
+
|
149 |
+
seqmatcher = dl.SequenceMatcher(None, s1, s2, autojunk=False)
|
150 |
+
res_orig, res_aug = [], []
|
151 |
+
for tag, a0, a1, b0, b1 in seqmatcher.get_opcodes():
|
152 |
+
if tag == "equal":
|
153 |
+
res_orig += s1[a0:a1]
|
154 |
+
res_aug += s2[b0:b1]
|
155 |
+
else:
|
156 |
+
res_orig += ["<span style='color:blue'> <em><b>"+" ".join(s1[a0:a1])+"</b></em></span>"]
|
157 |
+
res_aug += ["<span style='color:red'> <em><b>"+" ".join(s2[b0:b1])+"</b></em></span> "]
|
158 |
+
|
159 |
+
st.write("**<span style='font-size:larger'>The original text</span>**<br>(words are coloured in blue if they were selected for perturbation):", unsafe_allow_html=True)
|
160 |
+
st.write('<p style="padding: 1em">'+' '.join(res_orig)+'</p>', unsafe_allow_html=True)
|
161 |
+
|
162 |
+
|
163 |
+
st.write("**<span style='font-size:larger'>The perturbed text</span>**<br>(words are coloured in red if they appeared after perturbation):", unsafe_allow_html=True)
|
164 |
+
st.write('<p style="padding: 1em">'+' '.join(res_aug)+'</p>', unsafe_allow_html=True)
|