translatetopic / app.py
ManFrag's picture
added embeddings models
ab68b3c
raw
history blame
9.42 kB
import streamlit as st
import pandas as pd
from deep_translator import GoogleTranslator
from bertopic import BERTopic
from sentence_transformers import SentenceTransformer
from sklearn.feature_extraction.text import CountVectorizer
import base64
from io import BytesIO
import plotly.graph_objects as go
import plotly.subplots as sp
import plotly.express as px
def translate_feedback(feedback_df, column_name):
feedback_df["translated"] = "-" # Add a new column "translated" and initialize all rows with "-"
for i, feedback in enumerate(feedback_df[column_name]):
try:
translation = GoogleTranslator(source='auto', target='en').translate(feedback)
feedback_df.loc[i, "translated"] = translation # Store the translation in the "translated" column
except Exception as e:
feedback_df.loc[i, "translated"] = "-" # Store "-" in the "translated" column if an error occurs
feedback_df = feedback_df[feedback_df["translated"] != "-"] # Remove "-" rows
return feedback_df
@st.cache
def convert_df(df):
# IMPORTANT: Cache the conversion to prevent computation on every rerun
return df.to_csv().encode('utf-8')
def download_csv(df):
csv = df.to_csv(index=False)
b64 = base64.b64encode(csv.encode()).decode() # Encode the DataFrame as base64
href = f'<a href="data:file/csv;base64,{b64}" download="translated_feedback.csv">Download CSV file</a>'
return href
def topics_over_time(topic_model, dataframe, training_column):
timestamps = list(dataframe.day.values)
feedback_list = list(dataframe[training_column])
topics_over_time = topic_model.topics_over_time(feedback_list, timestamps, global_tuning=True, evolution_tuning=True)
f = topic_model.visualize_topics_over_time(topics_over_time, custom_labels=True)
f.update_layout(width=800,height=500)
return f
def area_over_time(topic_model, df, training_column, datetime_column):
df['Topic'] = topic_model.get_document_info(df[training_column])["Name"].values
df[datetime_column] = pd.to_datetime(df[datetime_column])
df['year'] = df[datetime_column].dt.year
df['month'] = df[datetime_column].dt.month
# Group the data by year, month, and topic
grouped = df.groupby(['year', 'month', 'Topic'])[training_column].count().reset_index()
# Normalize the document counts by the total document count for each month and topic
grouped['total_count'] = grouped.groupby(['year', 'month'])[training_column].transform('sum')
grouped['document_pct'] = grouped[training_column] / grouped['total_count'] * 100
# Pivot the data to create a table with months as rows, topics as columns, and document percentages as values
pivoted = pd.pivot_table(grouped, index=['year', 'month'], columns='Topic', values='document_pct', fill_value=0)
pivoted = pivoted.reset_index()
# Melt the data to create a long format with separate rows for each topic
melted = pd.melt(pivoted, id_vars=['year', 'month'], var_name='Topic', value_name='document_pct')
# Create the interactive plot using Plotly Express
fig = px.area(melted, x='month', y='document_pct', color='Topic', facet_col='year', facet_col_wrap=3,
title='Distribution of Documents by Topic and Month (Relative to 100%)',
labels={'month': 'Month', 'document_pct': 'Document Percentage', 'Topic': 'Topic', 'year': 'Year'},
hover_data={'month': False, 'document_pct': ':.2f'})
return fig
# Sidebar configuration
st.sidebar.title("Translation and Analysis App")
tab = st.sidebar.selectbox("Select Tab", ("Translate", "Analyse Feedback"))
if tab == "Translate":
st.title("Translate Feedback")
file = st.file_uploader("Upload CSV or Excel file", type=["csv", "xlsx"], accept_multiple_files=False)
if file is not None:
file.seek(0)
feedback_df = pd.read_csv(file, low_memory=False, on_bad_lines='skip', engine='c') if file.name.endswith(".csv") else pd.read_excel(file)
st.write('**Data Head:**')
st.write(feedback_df.head())
column_name = st.selectbox("Select Column", feedback_df.columns)
feedback_df = feedback_df.dropna(subset=[column_name])
feedback_df = feedback_df.reset_index(drop=True)
if st.button("Translate"):
translated_df = translate_feedback(feedback_df, column_name)
csv = convert_df(translated_df)
st.write('**Translated Data Head:**')
st.write(translated_df.head())
st.download_button(
label="Download data as CSV",
data=csv,
file_name='translated_data.csv',
mime='text/csv',
)
elif tab == "Analyse Feedback":
# Analyse Feedback tab code
st.title("Analyse Feedback")
file = st.file_uploader("Upload CSV or Excel file", type=["csv", "xlsx"])
if file is not None:
df = pd.read_csv(file, on_bad_lines='skip') if file.name.endswith(".csv") else pd.read_excel(file)
st.write('**Data Head:**')
st.write(df.head())
column_names = df.columns.tolist()
datetime_column = st.selectbox("Select Datetime Column", column_names + ["None"])
feedback_column = st.selectbox("Select Feedback Column", column_names)
model_select = st.selectbox(
"Select model to train:",
[
'all-mpnet-base-v2',
'all-distilroberta-v1',
'distiluse-base-multilingual-cased-v2',
'multi-qa-mpnet-base-dot-v1',
'multi-qa-distilbert-cos-v1',
'paraphrase-multilingual-mpnet-base-v2',
'BAAI/bge-small-en-v1.5',
'Cohere/Cohere-embed-english-v3.0'
]
)
if st.button("Train Model"):
if model_select is not None:
new_df = df.copy()
if datetime_column != "None":
new_df[datetime_column] = pd.to_datetime(new_df[datetime_column])
sentence_model = SentenceTransformer(model_select)
vectorizer_model = CountVectorizer(stop_words="english")
# Initialize a BERTopic model with the SentenceTransformer embeddings
my_model = BERTopic(
language="en",
calculate_probabilities=True,
verbose=True,
n_gram_range=(1, 3),
embedding_model=sentence_model,
vectorizer_model=vectorizer_model,
nr_topics = 15
)
# Preprocess the data by replacing missing values with empty strings
new_df[feedback_column] = new_df[feedback_column].fillna('')
new_df.reset_index(inplace = True,drop = True)
# Fit the BERTopic model on the dataframe
my_model.fit(new_df[feedback_column])
st.success("Model trained successfully")
# Store the trained model in session state
st.session_state.trained_model = my_model
st.session_state.new_df = new_df
st.session_state.feedback_colomn = feedback_column
st.session_state.datetime_column = datetime_column
if "trained_model" in st.session_state:
trained_model = st.session_state.trained_model
new_df = st.session_state.new_df
new_feedback_column = st.session_state.feedback_colomn
visualization_options = [
"Visualize documents",
"Topic Hierarchy",
"Barchart",
"Topics over time",
"Representative docs per topic"
]
selected_visualization = st.selectbox("Select Visualization", visualization_options)
if selected_visualization == "Barchart":
umap_fig = trained_model.visualize_barchart(n_words=5)
st.plotly_chart(umap_fig)
elif selected_visualization == "Visualize documents":
viz_doc = trained_model.visualize_documents(new_df[new_feedback_column])
st.plotly_chart(viz_doc)
elif selected_visualization == "Topic Hierarchy":
tsne_fig = trained_model.visualize_hierarchy(top_n_topics=20)
st.plotly_chart(tsne_fig)
elif selected_visualization == "Topics over time":
time_fig = area_over_time(trained_model, new_df, new_feedback_column, datetime_column)
st.plotly_chart(time_fig)
elif selected_visualization == "Representative docs per topic":
st.write(trained_model.get_representative_docs())
result = pd.merge(new_df[feedback_column],
trained_model.get_document_info(new_df[feedback_column]),
left_on=feedback_column,
right_on='Document',
how = 'left'
)
feedback_and_docs = convert_df(result)
st.download_button(
label="Download documents and topics",
data=feedback_and_docs,
file_name='document_info.csv',
mime='text/csv',
)