app-ai-ds-hec / pages /sentiment_analysis.py
laudavid's picture
fix page layout + poster recom movie
c548d40
raw
history blame
No virus
11.1 kB
import os
import re
import time
import streamlit as st
import matplotlib.pyplot as plt
import pandas as pd
import numpy as np
import altair as alt
import plotly.express as px
from pysentimiento import create_analyzer
from utils import load_data_pickle
st.set_page_config(layout="wide")
def clean_text(text):
pattern_punct = r"[^\w\s.',:/]"
pattern_date = r'\b\d{1,2}/\d{1,2}/\d{2,4}\b'
text = text.lower()
text = re.sub(pattern_date, '', text)
text = re.sub(pattern_punct, '', text)
text = text.replace("ggg","g")
text = text.replace(" "," ")
return text
@st.cache_data(ttl=3600, show_spinner=False)
def load_sa_model():
return create_analyzer(task="sentiment", lang="en")
st.markdown("# Sentiment Analysis")
st.markdown("### What is Sentiment Analysis ?")
st.info("""
Sentiment analysis is a **Natural Language Processing** (NLP) task that involves determining the sentiment or emotion expressed in a piece of text.
It has a wide range of use cases across various industries, as it helps organizations gain insights into the opinions, emotions, and attitudes expressed in text data.""")
st.markdown("Here is an example of Sentiment analysis used to analyze **Customer Satisfaction** for perfums.")
_, col, _ = st.columns([0.1,0.8,0.1])
with col:
st.image("images/sentiment_analysis.png") #, width=800)
st.markdown(" ")
st.markdown("""
Common applications of Natural Language Processing include:
- **Customer Feedback and Reviews** πŸ’―: Assessing reviews on products or services to understand customer satisfaction and identify areas for improvement.
- **Market Research** πŸ”: Analyzing survey responses or online forums to gauge public opinion on products, services, or emerging trends.
- **Financial Market Analysis** πŸ“‰: Monitoring financial news, reports, and social media to gauge investor sentiment and predict market trends.
- **Government and Public Policy** πŸ“£: Analyzing public opinion on government policies, initiatives, and political decisions to gauge public sentiment and inform decision-making.
""")
st.divider()
#sa_pages = ["Starbucks Customer Reviews (Text)", "Tiktok's US Congressional Hearing (Audio)"]
#st.markdown("### Select a use case ")
#use_case = st.selectbox("", sa_pages, label_visibility="collapsed")
st.markdown("### Customer Reviews πŸ“")
st.info("""In this use case, **sentiment analysis** is used to predict the **polarity** (negative, neutral, positive) of customer reviews.
You can try the application by using the provided starbucks customer reviews, or by writing your own.""")
st.markdown(" ")
_, col, _ = st.columns([0.25,0.5,0.25])
with col:
st.image("images/reviews.jpg")
st.markdown(" ")
# Load data
path_sa = "data/sa_data"
reviews_df = load_data_pickle(path_sa,"reviews_raw.pkl")
reviews_df.reset_index(drop=True, inplace=True)
reviews_df["Date"] = reviews_df["Date"].dt.date
reviews_df["Year"] = reviews_df["Year"].astype(int)
st.markdown("#### Predict polarity πŸ€”")
tab1_, tab2_ = st.tabs(["Starbucks reviews", "Write a review"])
with tab1_:
# FILTER DATA
st.markdown(" ")
col1, col2 = st.columns([0.2, 0.8], gap="medium")
with col1:
st.markdown("""<b>Filter reviews: </b> <br>
You can filter the dataset by Date, State or Rating""", unsafe_allow_html=True)
select_image_box = st.radio("",
["Filter by Date (Year)", "Filter by State", "Filter by Rating", "No filters"],
index=3, label_visibility="collapsed")
if select_image_box == "Filter by Date (Year)":
selected_date = st.multiselect("Date (Year)", reviews_df["Year"].unique(), default=reviews_df["Year"].unique()[0])
reviews_df = reviews_df.loc[reviews_df["Year"].isin(selected_date)]
if select_image_box == "Filter by State":
selected_state = st.multiselect("State", reviews_df["State"].unique(), default=reviews_df["State"].unique()[0])
reviews_df = reviews_df.loc[reviews_df["State"].isin(selected_state)]
if select_image_box == "Filter by Rating":
selected_rating = st.multiselect("Rating", sorted(list(reviews_df["Rating"].dropna().unique())),
default = sorted(list(reviews_df["Rating"].dropna().unique()))[0])
reviews_df = reviews_df.loc[reviews_df["Rating"].isin(selected_rating)]
if select_image_box == "No filters":
pass
#st.slider()
run_model1 = st.button("**Run the model**", type="primary", key="tab1")
st.info("The model has already been trained in this use case.")
with col2:
# VIEW DATA
st.markdown("""<b>View the reviews:</b> <br>
The dataset contains the location (State), date, rating, text and images (if provided) for each review.""",
unsafe_allow_html=True)
st.data_editor(
reviews_df.drop(columns=["Year"]),
column_config={"Image 1": st.column_config.ImageColumn("Image 1"),
"Image 2": st.column_config.ImageColumn("Image 2")},
hide_index=True)
######### SHOW RESULTS ########
if run_model1:
with st.spinner('Wait for it...'):
df_results = load_data_pickle(path_sa,"reviews_results.pkl")
df_results.reset_index(drop=True, inplace=True)
index_row = np.array(reviews_df.index)
df_results = df_results.iloc[index_row].reset_index(drop=True)
df_results["Review"] = reviews_df["Review"]
st.markdown(" ")
st.markdown("#### See the results β˜‘οΈ")
tab1, tab2, tab3 = st.tabs(["All results", "Results per state", "Results per year"])
with tab1: # Overall results (tab_1)
# get results df
df_results_tab1 = df_results[["ID","Review","Rating","Negative","Neutral","Positive","Result"]]
# warning message
df_warning = df_results_tab1["Result"].value_counts().to_frame().reset_index()
df_warning["Percentage"] = (100*df_warning["count"]/df_warning["count"].sum()).round(2)
perct_negative = df_warning.loc[df_warning["Result"]=="Negative","Percentage"].to_numpy()[0]
if perct_negative > 50:
st.error(f"**Negative reviews alert** ⚠️: The proportion of negative reviews is {perct_negative}% !")
# show dataframe results
st.data_editor(
df_results_tab1, #.loc[df_results_tab1["Customer ID"].isin(filter_customers)],
column_config={
"Negative": st.column_config.ProgressColumn(
"Negative πŸ‘Ž",
help="Negative score of the review",
format="%d%%",
min_value=0,
max_value=100),
"Neutral": st.column_config.ProgressColumn(
"Neutral βœ‹",
help="Neutral score of the review",
format="%d%%",
min_value=0,
max_value=100),
"Positive": st.column_config.ProgressColumn(
"Positive πŸ‘",
help="Positive score of the review",
format="%d%%",
min_value=0,
max_value=100)},
hide_index=True,
)
with tab2: # Results by state (tab_1)
avg_state = df_results[["State","Negative","Neutral","Positive"]].groupby(["State"]).mean().round()
avg_state = avg_state.reset_index().melt(id_vars="State", var_name="Sentiment", value_name="Score (%)")
chart_state = alt.Chart(avg_state, title="Review polarity per state").mark_bar().encode(
x=alt.X('Sentiment', axis=alt.Axis(title=None, labels=False, ticks=False)),
y=alt.Y('Score (%)', axis=alt.Axis(grid=False)),
color='Sentiment',
column=alt.Column('State', header=alt.Header(title=None, labelOrient='bottom'))
).configure_view(
stroke='transparent'
).interactive()
st.markdown(" ")
st.altair_chart(chart_state)
with tab3: # Results by year (tab_1)
avg_year = df_results[["Year","Negative","Neutral","Positive"]]
#avg_year["Year"] = avg_year["Year"].astype(str)
avg_year = avg_year.groupby(["Year"]).mean().round()
avg_year = avg_year.reset_index().melt(id_vars="Year", var_name="Sentiment", value_name="Score (%)")
chart_year = alt.Chart(avg_year, title="Evolution of review polarity").mark_area(opacity=0.5).encode(
x='Year',
y='Score (%)',
color='Sentiment',
).interactive()
st.markdown(" ")
st.altair_chart(chart_year, use_container_width=True)
# else:
# st.warning("You must select at least one review to run the model.")
#### WRITE YOUR OWN REVIEW #####""
with tab2_:
st.markdown("**Write your own review**")
txt_review = st.text_area(
"Write your review",
"I recently visited a local Starbucks, and unfortunately, my experience was far from satisfactory. "
"From the moment I stepped in, the atmosphere felt chaotic and disorganized. "
"The staff appeared overwhelmed, leading to a significant delay in receiving my order. "
"The quality of my drink further added to my disappointment. "
"The coffee tasted burnt, as if it had been sitting on the burner for far too long.",
label_visibility="collapsed"
)
run_model2 = st.button("**Run the model**", type="primary", key="tab2")
if run_model2:
with st.spinner('Wait for it...'):
#sentiment_analyzer = create_analyzer(task="sentiment", lang="en")
# Load model with cache
sentiment_analyzer = load_sa_model()
q = sentiment_analyzer.predict(txt_review)
df_review_user = pd.DataFrame({"Polarity":["Positive","Neutral","Negative"],
"Score":[q.probas['POS'], q.probas['NEU'], q.probas['NEG']]})
st.markdown(" ")
st.info(f"""Your review was **{int(q.probas['POS']*100)}%** positive, **{int(q.probas['NEU']*100)}%** neutral
and **{int(q.probas['NEG']*100)}%** negative.""")
fig = px.bar(df_review_user, x='Score', y='Polarity', color="Polarity", title='Sentiment analysis results', orientation="h")
st.plotly_chart(fig, use_container_width=True)