Updated app with code for deduplication
Browse files
app.py
CHANGED
@@ -4,11 +4,7 @@ import numpy as np
|
|
4 |
from model2vec import StaticModel
|
5 |
from reach import Reach
|
6 |
from tqdm import tqdm
|
7 |
-
import
|
8 |
-
|
9 |
-
def display_word_differences(x: str, y: str) -> str:
|
10 |
-
diff = difflib.ndiff(x.split(), y.split())
|
11 |
-
return " ".join([word for word in diff if word.startswith(('+', '-'))])
|
12 |
|
13 |
def deduplicate(embedding_matrix: np.ndarray, threshold: float, batch_size: int = 1024) -> tuple[np.ndarray, dict[int, int]]:
|
14 |
"""
|
@@ -24,11 +20,11 @@ def deduplicate(embedding_matrix: np.ndarray, threshold: float, batch_size: int
|
|
24 |
embedding_matrix,
|
25 |
threshold=threshold,
|
26 |
batch_size=batch_size,
|
27 |
-
show_progressbar=
|
28 |
)
|
29 |
|
30 |
# Process duplicates
|
31 |
-
for i, similar_items in enumerate(tqdm(results
|
32 |
if i not in deduplicated_indices:
|
33 |
continue # Skip already marked duplicates
|
34 |
|
@@ -58,11 +54,11 @@ def deduplicate_across_datasets(embedding_matrix_1: np.ndarray, embedding_matrix
|
|
58 |
embedding_matrix_2,
|
59 |
threshold=threshold,
|
60 |
batch_size=batch_size,
|
61 |
-
show_progressbar=
|
62 |
)
|
63 |
|
64 |
# Process duplicates
|
65 |
-
for i, similar_items in enumerate(tqdm(results
|
66 |
# Similar items are returned as (index, score), we are only interested in the index
|
67 |
similar_indices = [int(item[0]) for item in similar_items if item[1] >= threshold] # Keep those above the threshold
|
68 |
|
@@ -73,150 +69,154 @@ def deduplicate_across_datasets(embedding_matrix_1: np.ndarray, embedding_matrix
|
|
73 |
|
74 |
return duplicate_indices_in_test, duplicate_to_original_mapping
|
75 |
|
|
|
|
|
|
|
|
|
76 |
def perform_deduplication(
|
77 |
deduplication_type,
|
78 |
dataset1_name,
|
79 |
dataset1_split,
|
|
|
80 |
dataset2_name,
|
81 |
dataset2_split,
|
82 |
-
|
83 |
threshold
|
84 |
):
|
85 |
# Convert threshold to float
|
86 |
threshold = float(threshold)
|
87 |
-
|
88 |
-
|
89 |
-
|
90 |
-
|
91 |
-
|
92 |
-
|
93 |
-
|
94 |
-
|
95 |
-
|
96 |
-
|
97 |
-
|
98 |
-
|
99 |
-
|
100 |
-
|
101 |
-
model = StaticModel.from_pretrained("minishlab/M2V_base_output")
|
102 |
-
embedding_matrix = model.encode(texts, show_progressbar=False)
|
103 |
-
|
104 |
-
# Deduplicate
|
105 |
-
progress(0.5, desc="Performing deduplication...")
|
106 |
deduplicated_indices, duplicate_to_original_mapping = deduplicate(embedding_matrix, threshold)
|
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 |
-
|
149 |
-
|
150 |
-
|
151 |
-
|
152 |
-
progress(0.5, desc="Computing embeddings for Dataset 2...")
|
153 |
-
embedding_matrix2 = model.encode(texts2, show_progressbar=False)
|
154 |
-
|
155 |
-
# Deduplicate across datasets
|
156 |
-
progress(0.7, desc="Performing cross-dataset deduplication...")
|
157 |
duplicate_indices_in_ds2, duplicate_to_original_mapping = deduplicate_across_datasets(embedding_matrix1, embedding_matrix2, threshold)
|
158 |
-
|
159 |
-
|
160 |
-
|
161 |
-
|
162 |
-
|
163 |
-
|
164 |
-
|
165 |
-
|
166 |
-
|
167 |
-
|
168 |
-
|
169 |
-
|
170 |
-
|
171 |
-
|
172 |
-
|
173 |
-
|
174 |
-
|
175 |
-
|
176 |
-
|
177 |
-
|
178 |
-
|
179 |
-
|
180 |
-
|
181 |
-
|
182 |
-
|
183 |
-
return result_text
|
184 |
|
185 |
with gr.Blocks() as demo:
|
186 |
gr.Markdown("# Semantic Deduplication")
|
187 |
-
|
188 |
deduplication_type = gr.Radio(choices=["Single dataset", "Cross-dataset"], label="Deduplication Type", value="Single dataset")
|
189 |
-
|
190 |
-
with gr.
|
191 |
-
|
192 |
-
|
193 |
-
|
194 |
-
|
195 |
-
|
196 |
-
|
197 |
-
|
198 |
-
|
199 |
-
|
200 |
-
|
|
|
|
|
201 |
threshold = gr.Slider(minimum=0.0, maximum=1.0, value=0.8, label="Similarity Threshold")
|
202 |
-
|
203 |
compute_button = gr.Button("Compute")
|
204 |
-
|
205 |
output = gr.Markdown()
|
206 |
-
|
207 |
-
# Function to update the visibility of
|
208 |
-
def update_visibility(
|
209 |
-
if
|
210 |
-
return {
|
211 |
else:
|
212 |
-
return {
|
|
|
|
|
213 |
|
214 |
-
deduplication_type.change(update_visibility, inputs=deduplication_type, outputs=[dataset2_row])
|
215 |
-
|
216 |
compute_button.click(
|
217 |
fn=perform_deduplication,
|
218 |
-
inputs=[
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
219 |
outputs=output
|
220 |
)
|
221 |
-
|
222 |
demo.launch()
|
|
|
4 |
from model2vec import StaticModel
|
5 |
from reach import Reach
|
6 |
from tqdm import tqdm
|
7 |
+
from difflib import ndiff
|
|
|
|
|
|
|
|
|
8 |
|
9 |
def deduplicate(embedding_matrix: np.ndarray, threshold: float, batch_size: int = 1024) -> tuple[np.ndarray, dict[int, int]]:
|
10 |
"""
|
|
|
20 |
embedding_matrix,
|
21 |
threshold=threshold,
|
22 |
batch_size=batch_size,
|
23 |
+
show_progressbar=True
|
24 |
)
|
25 |
|
26 |
# Process duplicates
|
27 |
+
for i, similar_items in enumerate(tqdm(results)):
|
28 |
if i not in deduplicated_indices:
|
29 |
continue # Skip already marked duplicates
|
30 |
|
|
|
54 |
embedding_matrix_2,
|
55 |
threshold=threshold,
|
56 |
batch_size=batch_size,
|
57 |
+
show_progressbar=True
|
58 |
)
|
59 |
|
60 |
# Process duplicates
|
61 |
+
for i, similar_items in enumerate(tqdm(results)):
|
62 |
# Similar items are returned as (index, score), we are only interested in the index
|
63 |
similar_indices = [int(item[0]) for item in similar_items if item[1] >= threshold] # Keep those above the threshold
|
64 |
|
|
|
69 |
|
70 |
return duplicate_indices_in_test, duplicate_to_original_mapping
|
71 |
|
72 |
+
def display_word_differences(x: str, y: str) -> str:
|
73 |
+
diff = ndiff(x.split(), y.split())
|
74 |
+
return " ".join([f"{word}" for word in diff if word.startswith(('+', '-'))])
|
75 |
+
|
76 |
def perform_deduplication(
|
77 |
deduplication_type,
|
78 |
dataset1_name,
|
79 |
dataset1_split,
|
80 |
+
dataset1_text_column,
|
81 |
dataset2_name,
|
82 |
dataset2_split,
|
83 |
+
dataset2_text_column,
|
84 |
threshold
|
85 |
):
|
86 |
# Convert threshold to float
|
87 |
threshold = float(threshold)
|
88 |
+
|
89 |
+
if deduplication_type == "Single dataset":
|
90 |
+
# Load the dataset
|
91 |
+
ds = load_dataset(dataset1_name, split=dataset1_split)
|
92 |
+
|
93 |
+
# Extract texts
|
94 |
+
texts = [example[dataset1_text_column] for example in ds]
|
95 |
+
|
96 |
+
# Compute embeddings
|
97 |
+
model = StaticModel.from_pretrained("minishlab/M2V_base_output")
|
98 |
+
embedding_matrix = model.encode(texts, show_progressbar=True)
|
99 |
+
|
100 |
+
# Deduplicate
|
101 |
+
with gr.Progress(track_tqdm=True):
|
|
|
|
|
|
|
|
|
|
|
102 |
deduplicated_indices, duplicate_to_original_mapping = deduplicate(embedding_matrix, threshold)
|
103 |
+
|
104 |
+
# Prepare the results
|
105 |
+
num_duplicates = len(duplicate_to_original_mapping)
|
106 |
+
num_total = len(texts)
|
107 |
+
num_deduplicated = len(deduplicated_indices)
|
108 |
+
|
109 |
+
result_text = f"**Total documents:** {num_total}\n"
|
110 |
+
result_text += f"**Number of duplicates found:** {num_duplicates}\n"
|
111 |
+
result_text += f"**Number of unique documents after deduplication:** {num_deduplicated}\n\n"
|
112 |
+
|
113 |
+
# Show deduplicated examples
|
114 |
+
result_text += "**Examples of duplicates found:**\n\n"
|
115 |
+
num_examples = min(5, num_duplicates)
|
116 |
+
examples_shown = 0
|
117 |
+
for duplicate_idx, original_idx in duplicate_to_original_mapping.items():
|
118 |
+
if examples_shown >= num_examples:
|
119 |
+
break
|
120 |
+
original_text = texts[original_idx]
|
121 |
+
duplicate_text = texts[duplicate_idx]
|
122 |
+
differences = display_word_differences(original_text, duplicate_text)
|
123 |
+
result_text += f"**Original text:**\n{original_text}\n\n"
|
124 |
+
result_text += f"**Duplicate text:**\n{duplicate_text}\n\n"
|
125 |
+
result_text += f"**Differences:**\n{differences}\n"
|
126 |
+
result_text += "-" * 50 + "\n\n"
|
127 |
+
examples_shown += 1
|
128 |
+
|
129 |
+
return result_text
|
130 |
+
|
131 |
+
elif deduplication_type == "Cross-dataset":
|
132 |
+
# Load datasets
|
133 |
+
ds1 = load_dataset(dataset1_name, split=dataset1_split)
|
134 |
+
ds2 = load_dataset(dataset2_name, split=dataset2_split)
|
135 |
+
|
136 |
+
# Extract texts
|
137 |
+
texts1 = [example[dataset1_text_column] for example in ds1]
|
138 |
+
texts2 = [example[dataset2_text_column] for example in ds2]
|
139 |
+
|
140 |
+
# Compute embeddings
|
141 |
+
model = StaticModel.from_pretrained("minishlab/M2V_base_output")
|
142 |
+
embedding_matrix1 = model.encode(texts1, show_progressbar=True)
|
143 |
+
embedding_matrix2 = model.encode(texts2, show_progressbar=True)
|
144 |
+
|
145 |
+
# Deduplicate across datasets
|
146 |
+
with gr.Progress(track_tqdm=True):
|
|
|
|
|
|
|
|
|
|
|
|
|
147 |
duplicate_indices_in_ds2, duplicate_to_original_mapping = deduplicate_across_datasets(embedding_matrix1, embedding_matrix2, threshold)
|
148 |
+
|
149 |
+
num_duplicates = len(duplicate_indices_in_ds2)
|
150 |
+
num_total_ds2 = len(texts2)
|
151 |
+
num_unique_ds2 = num_total_ds2 - num_duplicates
|
152 |
+
|
153 |
+
result_text = f"**Total documents in {dataset2_name}/{dataset2_split}:** {num_total_ds2}\n"
|
154 |
+
result_text += f"**Number of duplicates found in {dataset2_name}/{dataset2_split}:** {num_duplicates}\n"
|
155 |
+
result_text += f"**Number of unique documents in {dataset2_name}/{dataset2_split} after deduplication:** {num_unique_ds2}\n\n"
|
156 |
+
|
157 |
+
# Show deduplicated examples
|
158 |
+
result_text += "**Examples of duplicates found in Dataset 2:**\n\n"
|
159 |
+
num_examples = min(5, num_duplicates)
|
160 |
+
examples_shown = 0
|
161 |
+
for duplicate_idx in duplicate_indices_in_ds2[:num_examples]:
|
162 |
+
original_idx = duplicate_to_original_mapping[duplicate_idx]
|
163 |
+
original_text = texts1[original_idx]
|
164 |
+
duplicate_text = texts2[duplicate_idx]
|
165 |
+
differences = display_word_differences(original_text, duplicate_text)
|
166 |
+
result_text += f"**Original text (Dataset 1):**\n{original_text}\n\n"
|
167 |
+
result_text += f"**Duplicate text (Dataset 2):**\n{duplicate_text}\n\n"
|
168 |
+
result_text += f"**Differences:**\n{differences}\n"
|
169 |
+
result_text += "-" * 50 + "\n\n"
|
170 |
+
examples_shown += 1
|
171 |
+
|
172 |
+
return result_text
|
|
|
173 |
|
174 |
with gr.Blocks() as demo:
|
175 |
gr.Markdown("# Semantic Deduplication")
|
176 |
+
|
177 |
deduplication_type = gr.Radio(choices=["Single dataset", "Cross-dataset"], label="Deduplication Type", value="Single dataset")
|
178 |
+
|
179 |
+
with gr.Tab("Dataset 1"):
|
180 |
+
with gr.Row():
|
181 |
+
dataset1_name = gr.Textbox(value="ag_news", label="Dataset Name")
|
182 |
+
dataset1_split = gr.Textbox(value="train", label="Split")
|
183 |
+
dataset1_text_column = gr.Textbox(value="text", label="Text Column Name")
|
184 |
+
|
185 |
+
dataset2_tab = gr.Tab("Dataset 2", visible=False)
|
186 |
+
with dataset2_tab:
|
187 |
+
with gr.Row():
|
188 |
+
dataset2_name = gr.Textbox(value="ag_news", label="Dataset Name")
|
189 |
+
dataset2_split = gr.Textbox(value="test", label="Split")
|
190 |
+
dataset2_text_column = gr.Textbox(value="text", label="Text Column Name")
|
191 |
+
|
192 |
threshold = gr.Slider(minimum=0.0, maximum=1.0, value=0.8, label="Similarity Threshold")
|
193 |
+
|
194 |
compute_button = gr.Button("Compute")
|
195 |
+
|
196 |
output = gr.Markdown()
|
197 |
+
|
198 |
+
# Function to update the visibility of dataset2_tab
|
199 |
+
def update_visibility(deduplication_type):
|
200 |
+
if deduplication_type == "Cross-dataset":
|
201 |
+
return {dataset2_tab: gr.update(visible=True)}
|
202 |
else:
|
203 |
+
return {dataset2_tab: gr.update(visible=False)}
|
204 |
+
|
205 |
+
deduplication_type.change(update_visibility, inputs=deduplication_type, outputs=[dataset2_tab])
|
206 |
|
|
|
|
|
207 |
compute_button.click(
|
208 |
fn=perform_deduplication,
|
209 |
+
inputs=[
|
210 |
+
deduplication_type,
|
211 |
+
dataset1_name,
|
212 |
+
dataset1_split,
|
213 |
+
dataset1_text_column,
|
214 |
+
dataset2_name,
|
215 |
+
dataset2_split,
|
216 |
+
dataset2_text_column,
|
217 |
+
threshold
|
218 |
+
],
|
219 |
outputs=output
|
220 |
)
|
221 |
+
|
222 |
demo.launch()
|