clusters / app.py
pentarosarium's picture
1.23 print debug
25558c7
import streamlit as st
import pandas as pd
import numpy as np
from transformers import AutoTokenizer, AutoModel
import torch
from datetime import datetime
import io
import base64
from typing import Dict, List, Set, Tuple
from rapidfuzz import fuzz, process
from collections import defaultdict
from tqdm import tqdm
import spacy
import torch.nn.functional as F
class NewsProcessor:
def __init__(self, similarity_threshold=0.75, time_threshold=24):
try:
self.nlp = spacy.load("ru_core_news_sm")
except:
self.nlp = spacy.load("en_core_web_sm")
import pymorphy2
self.morph = pymorphy2.MorphAnalyzer()
self.tokenizer = AutoTokenizer.from_pretrained('sentence-transformers/paraphrase-multilingual-mpnet-base-v2')
self.model = AutoModel.from_pretrained('sentence-transformers/paraphrase-multilingual-mpnet-base-v2')
self.similarity_threshold = similarity_threshold
self.time_threshold = time_threshold
def mean_pooling(self, model_output, attention_mask):
token_embeddings = model_output[0]
input_mask_expanded = attention_mask.unsqueeze(-1).expand(token_embeddings.size()).float()
return torch.sum(token_embeddings * input_mask_expanded, 1) / torch.clamp(input_mask_expanded.sum(1), min=1e-9)
def encode_text(self, text):
# Convert text to string and handle NaN values
if pd.isna(text):
text = ""
else:
text = str(text)
encoded_input = self.tokenizer(text, padding=True, truncation=True, max_length=512, return_tensors='pt')
with torch.no_grad():
model_output = self.model(**encoded_input)
sentence_embeddings = self.mean_pooling(model_output, encoded_input['attention_mask'])
return F.normalize(sentence_embeddings[0], p=2, dim=0).numpy()
def get_company_variants(self, company_name: str) -> Set[str]:
"""Generate morphological variants of company name."""
if pd.isna(company_name):
return set()
# Clean the company name
name = str(company_name).strip('"\'').strip()
name = name.split(',')[0].strip() # Take first part before comma
variants = set()
variants.add(name.lower())
# Split into words and get significant parts
words = [w for w in name.split() if len(w) >= 3]
# Generate morphological variants for each significant word
for word in words:
parsed = self.morph.parse(word)[0]
lexeme = parsed.lexeme
variants.update(v.word.lower() for v in lexeme)
# Add combinations of consecutive words
if len(words) > 1:
for i in range(len(words)-1):
variants.add(f"{words[i]} {words[i+1]}".lower())
return variants
def is_company_main_subject(self, title: str, text: str, company_name: str, threshold_score: float = 0.5) -> Tuple[bool, float]:
"""
Enhanced company subject detection using title and text with Russian language support.
Returns (is_main_subject, relevance_score).
"""
if pd.isna(text) or pd.isna(company_name):
return False, 0.0
# Ensure we have strings
title = str(title) if not pd.isna(title) else ""
text = str(text) if not pd.isna(text) else ""
# Get company name variants
company_variants = self.get_company_variants(company_name)
if not company_variants:
return False, 0.0
# Initialize scoring components
title_score = 0.0
first_para_score = 0.0
subject_score = 0.0
frequency_score = 0.0
# Process title (weight: 0.4)
title_doc = self.nlp(title.lower())
title_text = title_doc.text
for variant in company_variants:
if variant in title_text:
title_score = 0.4
# Check if company is subject in title
for token in title_doc:
if variant in token.text and token.dep_ in ['nsubj', 'nsubjpass']:
title_score = 0.4
break
break
# Process main text
doc = self.nlp(text.lower())
paragraphs = [p.strip() for p in text.split('\n') if p.strip()]
first_para = paragraphs[0] if paragraphs else ""
# Check first paragraph (weight: 0.2)
for variant in company_variants:
if variant in first_para.lower():
first_para_score = 0.2
break
# Analyze subject position and frequency
company_mentions = 0
subject_mentions = 0
other_company_indicators = {
'компания', 'корпорация', 'фирма', 'банк', 'группа', 'холдинг',
'организация', 'предприятие', 'производитель', 'ао', 'оао', 'пао', 'нк', 'гк',
'ооо', 'лк', 'фк', 'акб', 'ук', 'зао', 'ак'
}
other_companies = 0
# Analyze each sentence
for sent in doc.sents:
sent_text = sent.text.lower()
# Count company mentions and subject positions
company_in_sent = False
for variant in company_variants:
if variant in sent_text:
company_mentions += 1
company_in_sent = True
# Check subject position
for token in sent:
if variant in token.text and token.dep_ in ['nsubj', 'nsubjpass']:
subject_mentions += 1
# Count other company mentions
if company_in_sent:
continue
for indicator in other_company_indicators:
if indicator in sent_text:
other_companies += 1
break
# Calculate subject score (weight: 0.2)
subject_score = min(0.2, (subject_mentions / max(1, company_mentions)) * 0.2)
# Calculate frequency score (weight: 0.2)
if company_mentions > 0:
company_ratio = company_mentions / (company_mentions + other_companies + 1)
frequency_score = min(0.2, company_ratio * 0.2)
# Calculate final score
final_score = title_score + first_para_score + subject_score + frequency_score
# Apply penalties
if other_companies > 5: # Too many other companies mentioned
final_score *= 0.5
# Check if the company is just part of a list
list_indicators = {'среди', 'включая', 'такие как', 'в том числе', 'и другие', 'а также'}
for indicator in list_indicators:
if indicator in text.lower():
final_score *= 0.7
return final_score >= threshold_score, final_score
def process_news(self, df: pd.DataFrame, progress_bar=None):
# Ensure the DataFrame is not empty
if df.empty:
return pd.DataFrame(columns=['cluster_id', 'datetime', 'company', 'relevance_score', 'text', 'cluster_size'])
df = df.copy() # Make a copy to preserve original indices
clusters = []
processed = set()
for idx in df.index: # Iterate over original indices
if idx in processed:
continue
row1 = df.loc[idx]
cluster = [idx] # Store original index
processed.add(idx)
if not pd.isna(row1['text']):
text1_embedding = self.encode_text(row1['text'])
if progress_bar:
progress_bar.progress(len(processed) / len(df))
for other_idx in df.index: # Iterate over original indices
if other_idx in processed:
continue
row2 = df.loc[other_idx]
if pd.isna(row2['text']):
continue
time_diff = pd.to_datetime(row1['datetime']) - pd.to_datetime(row2['datetime'])
if abs(time_diff.total_seconds() / 3600) > self.time_threshold:
continue
text2_embedding = self.encode_text(row2['text'])
similarity = np.dot(text1_embedding, text2_embedding)
if similarity >= self.similarity_threshold:
cluster.append(other_idx)
processed.add(other_idx)
clusters.append(cluster)
# Create result DataFrame preserving original indices
result_data = []
for cluster_id, cluster_indices in enumerate(clusters, 1):
cluster_rows = df.loc[cluster_indices]
for idx in cluster_indices:
result_data.append({
'cluster_id': cluster_id,
'datetime': df.loc[idx, 'datetime'],
'company': df.loc[idx, 'company'],
'text': df.loc[idx, 'text'],
'cluster_size': len(cluster_indices)
})
result_df = pd.DataFrame(result_data, index=sum(clusters, [])) # Use original indices
return result_df
class NewsDeduplicator:
def __init__(self, fuzzy_threshold=85):
self.fuzzy_threshold = fuzzy_threshold
def deduplicate(self, df: pd.DataFrame, progress_bar=None) -> pd.DataFrame:
seen_texts: List[str] = []
text_to_companies: Dict[str, Set[str]] = defaultdict(set)
indices_to_keep: Set[int] = set()
for idx, row in tqdm(df.iterrows(), total=len(df)):
text = str(row['text']) if not pd.isna(row['text']) else ""
company = str(row['company']) if not pd.isna(row['company']) else ""
if not text:
indices_to_keep.add(idx)
continue
if seen_texts:
result = process.extractOne(
text,
seen_texts,
scorer=fuzz.ratio,
score_cutoff=self.fuzzy_threshold
)
match = result[0] if result else None
else:
match = None
if match:
text_to_companies[match].add(company)
else:
seen_texts.append(text)
text_to_companies[text].add(company)
indices_to_keep.add(idx)
if progress_bar:
progress_bar.progress((idx + 1) / len(df))
dedup_df = df.iloc[list(indices_to_keep)].copy()
for idx in indices_to_keep:
text = str(df.iloc[idx]['text'])
companies = sorted(text_to_companies[text])
dedup_df.at[idx, 'company'] = ' | '.join(companies)
dedup_df.at[idx, 'company_count'] = len(companies)
dedup_df.at[idx, 'duplicate_count'] = len(text_to_companies[text])
return dedup_df.sort_values('datetime')
def create_download_link(df: pd.DataFrame, filename: str) -> str:
excel_buffer = io.BytesIO()
with pd.ExcelWriter(excel_buffer, engine='openpyxl') as writer:
df.to_excel(writer, index=False)
excel_buffer.seek(0)
b64 = base64.b64encode(excel_buffer.read()).decode()
return f'<a href="data:application/vnd.openxmlformats-officedocument.spreadsheetml.sheet;base64,{b64}" download="{filename}">Download {filename}</a>'
def main():
st.title("кластеризуем новости v.1.23 + print debug")
st.write("Upload Excel file with columns: company, datetime, text")
uploaded_file = st.file_uploader("Choose Excel file", type=['xlsx'])
if uploaded_file:
try:
# Read all columns from original sheet
df_original = pd.read_excel(uploaded_file, sheet_name='Публикации')
st.write("Available columns:", df_original.columns.tolist())
# Create working copy with required columns
df = df_original.copy()
text_column = df_original.columns[6]
title_column = df_original.columns[5]
datetime_column = df_original.columns[3]
company_column = df_original.columns[0]
df = df_original[[company_column, datetime_column, title_column, text_column]].copy()
df.columns = ['company', 'datetime', 'title', 'text']
st.success(f'Loaded {len(df)} records')
st.dataframe(df.head())
col1, col2 = st.columns(2)
with col1:
fuzzy_threshold = st.slider("Fuzzy Match Threshold", 30, 100, 50)
with col2:
similarity_threshold = st.slider("Similarity Threshold", 0.5, 1.0, 0.75)
time_threshold = st.slider("Time Threshold (hours)", 1, 72, 24)
if st.button("Process News"):
try:
progress_bar = st.progress(0)
# Step 1: Deduplicate
deduplicator = NewsDeduplicator(fuzzy_threshold)
dedup_df = deduplicator.deduplicate(df, progress_bar)
# Preserve all columns from original DataFrame in dedup_df_full
dedup_df_full = df_original.loc[dedup_df.index].copy()
st.write("\nDeduplication Results:")
st.write(f"Original indices: {df.index.tolist()}")
st.write(f"Dedup indices: {dedup_df.index.tolist()}")
st.write(f"Sample from dedup_df:")
st.write(dedup_df[['company', 'text']].head())
# Step 2: Cluster deduplicated news
processor = NewsProcessor(similarity_threshold, time_threshold)
result_df = processor.process_news(dedup_df, progress_bar)
st.write("\nClustering Results:")
st.write(f"Result df indices: {result_df.index.tolist()}")
# Display cluster information
if len(result_df) > 0:
st.write("\nCluster Details:")
for cluster_id in result_df['cluster_id'].unique():
cluster_mask = result_df['cluster_id'] == cluster_id
if sum(cluster_mask) > 1: # Only show multi-item clusters
cluster_indices = result_df[cluster_mask].index.tolist()
st.write(f"\nCluster {cluster_id}:")
st.write(f"Indices: {cluster_indices}")
# Show texts for verification
for idx in cluster_indices:
text_length = len(str(dedup_df.loc[idx, 'text']))
st.write(f"Index {idx} - Length {text_length}:")
st.write(str(dedup_df.loc[idx, 'text'])[:100] + '...')
# Process clusters for deletion
indices_to_delete = set()
if len(result_df) > 0:
for cluster_id in result_df['cluster_id'].unique():
cluster_mask = result_df['cluster_id'] == cluster_id
if sum(cluster_mask) > 1:
cluster_indices = result_df[cluster_mask].index.tolist()
text_lengths = dedup_df.loc[cluster_indices, 'text'].fillna('').str.len()
longest_text_idx = text_lengths.idxmax()
indices_to_delete.update(set(cluster_indices) - {longest_text_idx})
st.write("\nDeletion Summary:")
st.write(f"Indices to delete: {sorted(list(indices_to_delete))}")
# Create final DataFrame
declustered_df = dedup_df_full.copy()
if indices_to_delete:
declustered_df = declustered_df.drop(index=list(indices_to_delete))
st.write(f"Final indices kept: {sorted(declustered_df.index.tolist())}")
# Print statistics
st.success(f"""
Processing results:
- Original rows: {len(df_original)}
- After deduplication: {len(dedup_df_full)}
- Multi-item clusters found: {len(result_df[result_df['cluster_size'] > 1]['cluster_id'].unique()) if len(result_df) > 0 else 0}
- Rows removed from clusters: {len(indices_to_delete)}
- Final rows kept: {len(declustered_df)}
""")
# Download buttons for all results
st.subheader("Download Results")
st.markdown(create_download_link(dedup_df_full, "deduplicated_news.xlsx"), unsafe_allow_html=True)
st.markdown(create_download_link(result_df, "clustered_news.xlsx"), unsafe_allow_html=True)
st.markdown(create_download_link(declustered_df, "declustered_news.xlsx"), unsafe_allow_html=True)
# Show clusters info
if len(result_df) > 0:
st.subheader("Largest Clusters")
largest_clusters = result_df[result_df['cluster_size'] > 1].sort_values(
['cluster_size', 'cluster_id', 'datetime'],
ascending=[False, True, True]
)
st.dataframe(largest_clusters)
except Exception as e:
st.error(f"Error: {str(e)}")
import traceback
st.error(traceback.format_exc())
finally:
progress_bar.empty()
except Exception as e:
st.error(f"Error reading file: {str(e)}")
import traceback
st.error(traceback.format_exc())
if __name__ == "__main__":
main()