app.py

import streamlit as st
from transformers import AutoModelForSequenceClassification, AutoTokenizer, pipeline
import spacy
from tika import parser
import requests
import pandas as pd

# Loading spaCy model outside the streamlit cache
nlp = spacy.load("en_core_web_sm")

@st.cache(allow_output_mutation=True)
def load_environmental_model():
    name_env = "ESGBERT/EnvironmentalBERT-environmental"
    tokenizer_env = AutoTokenizer.from_pretrained(name_env)
    model_env = AutoModelForSequenceClassification.from_pretrained(name_env)
    return pipeline("text-classification", model=model_env, tokenizer=tokenizer_env)

@st.cache(allow_output_mutation=True)
def load_social_model():
    name_soc = "ESGBERT/SocialBERT-social"
    tokenizer_soc = AutoTokenizer.from_pretrained(name_soc)
    model_soc = AutoModelForSequenceClassification.from_pretrained(name_soc)
    return pipeline("text-classification", model=model_soc, tokenizer=tokenizer_soc)

@st.cache(allow_output_mutation=True)
def load_governance_model():
    name_gov = "ESGBERT/GovernanceBERT-governance"
    tokenizer_gov = AutoTokenizer.from_pretrained(name_gov)
    model_gov = AutoModelForSequenceClassification.from_pretrained(name_gov)
    return pipeline("text-classification", model=model_gov, tokenizer=tokenizer_gov)

@st.cache(allow_output_mutation=True)
def load_sentiment_model():
    model_name = "climatebert/distilroberta-base-climate-sentiment"
    model = AutoModelForSequenceClassification.from_pretrained(model_name)
    tokenizer = AutoTokenizer.from_pretrained(model_name, max_len=512)
    return pipeline("text-classification", model=model, tokenizer=tokenizer)

# Streamlit App
st.title("ESGBERT Text Classification App")

# Get report URL from user input
url = st.text_input("Enter the URL of the report (PDF):")

# Model selection dropdown
selected_model = st.selectbox("Select Model", ["Environmental Model", "Social Model", "Governance Model", "Sentiment Model"])

if url:
    # Download PDF content from the URL
    response = requests.get(url, stream=True)

    if response.status_code == 200:
        # Parse PDF and extract text
        raw_text = parser.from_buffer(response.content)['content']
        
        # Extract sentences using spaCy
        doc = nlp(raw_text)
        sentences = [sent.text for sent in doc.sents]

        # Filtering and preprocessing sentences
        sequences = list(map(str, sentences))
        sentences = [x.replace("\n", "") for x in sequences]
        sentences = [x for x in sentences if x != ""]
        sentences = [x for x in sentences if x[0].isupper()]
        sub_sentences = sentences[:100]  # Takes around 20 seconds

        # Classification using different models based on user selection
        if selected_model == "Environmental Model":
            pipe_model = load_environmental_model()
        elif selected_model == "Social Model":
            pipe_model = load_social_model()
        elif selected_model == "Governance Model":
            pipe_model = load_governance_model()
        else:
            pipe_model = load_sentiment_model()

        # Get predictions for the selected model
        model_results = pipe_model(sub_sentences, padding=True, truncation=True)
        model_labels = [x["label"] for x in model_results]

        # Display count of sentences labeled as the selected model
        st.subheader(f"{selected_model} Sentences Count")
        st.write(pd.DataFrame({"sentence": sub_sentences, selected_model: model_labels}).groupby(selected_model).count())

    else:
        st.error("Error fetching PDF content from the provided URL. Please check the URL and try again.")