app.py

import streamlit as st
import torch
from transformers import AutoTokenizer, AutoModelForSequenceClassification
from collections import defaultdict
import fitz  # PyMuPDF for PDF reading
import re
import os
import spacy  # Replace NLTK with spaCy for sentence tokenization

# Load spaCy model for sentence tokenization
nlp = spacy.load("en_core_web_sm")

# Streamlit App Configuration
st.set_page_config(page_title="📊 Financial Report Sentiment Analyzer", layout="wide")
st.title("📊 Financial Report Sentiment Analyzer")

st.markdown("""
### What is FinBERT?
**FinBERT** is a language model fine-tuned for financial text analysis. It classifies sentiment as **Positive, Neutral, or Negative** for key financial aspects:
1. **Assets** – What the company owns
2. **Liabilities** – What the company owes
3. **Equity** – Net worth (Assets - Liabilities)
---
""")

# File Upload
uploaded_file = st.file_uploader("📂 Upload Financial Report (.pdf or .txt)", type=["pdf", "txt"])

# ✅ Custom CSS for Better Report Preview
st.markdown("""
<style>
.report-preview {
    border: 1px solid #ccc;
    padding: 10px;
    max-height: 300px;
    overflow-y: scroll;
    background-color: #f9f9f9;
    color: #333 !important;
    white-space: pre-wrap;
    line-height: 1.6;
    font-family: Arial, sans-serif;
}
</style>
""", unsafe_allow_html=True)

# ✅ Load FinBERT Model (Optimized with Streamlit Caching)
@st.cache_resource
def load_model():
    tokenizer = AutoTokenizer.from_pretrained("yiyanghkust/finbert-tone")
    model = AutoModelForSequenceClassification.from_pretrained("yiyanghkust/finbert-tone")
    return tokenizer, model

tokenizer, model = load_model()
label_mapping = {0: 'Positive', 1: 'Negative', 2: 'Neutral'}

# ✅ Extract Text from Uploaded File
def extract_text(file):
    try:
        if file.name.endswith('.pdf'):
            with fitz.open(stream=file.read(), filetype="pdf") as doc:
                return "\n".join([page.get_text() for page in doc])
        else:
            return file.read().decode('utf-8')
    except Exception as e:
        st.error(f"❌ Error reading file: {e}")
        return ""

if uploaded_file:
    report_text = extract_text(uploaded_file)
    st.write("### 📄 Uploaded Report Preview:")
    st.markdown(f"<div class='report-preview'>{report_text[:5000]}</div>", unsafe_allow_html=True)

    # ✅ Sentiment Analysis Function
    def analyze_sentiment(sentence):
        inputs = tokenizer(sentence, return_tensors="pt", truncation=True, max_length=512)
        with torch.no_grad():
            outputs = model(**inputs)
        probs = torch.nn.functional.softmax(outputs.logits, dim=-1)
        label_idx = torch.argmax(probs, dim=1).item()
        return label_mapping[label_idx], probs.tolist()[0]

    # ✅ Extract Sentences Matching Financial Keywords (using spaCy)
    def extract_sentences(text, keywords):
        try:
            doc = nlp(text)
            sentences = [sent.text for sent in doc.sents]  # Use spaCy for sentence tokenization
            pattern = re.compile(r'\b(' + '|'.join(map(re.escape, keywords)) + r')\b', re.IGNORECASE)
            return [s for s in sentences if pattern.search(s)]
        except Exception as e:
            st.error(f"❌ Error in sentence tokenization: {e}")
            return []

    # ✅ Analyze Sentiment for a Specific Financial Category
    def analyze_category(text, category_name, keywords):
        sentences = extract_sentences(text, keywords)
        if not sentences:
            st.warning(f"⚠️ No relevant sentences found for {category_name}")
            return None, []
        
        sentiment_scores = defaultdict(int)
        negative_sentences = []

        for sentence in sentences:
            label, probs = analyze_sentiment(sentence)
            sentiment_scores[label] += 1
            if label == 'Negative':
                negative_sentences.append((sentence, probs))

        total = sum(sentiment_scores.values())
        sentiment_percentages = {
            'Positive': (sentiment_scores.get('Positive', 0) / total) * 100 if total else 0,
            'Negative': (sentiment_scores.get('Negative', 0) / total) * 100 if total else 0,
            'Neutral': (sentiment_scores.get('Neutral', 0) / total) * 100 if total else 0
        }
        return sentiment_percentages, negative_sentences

    # ✅ Financial Categories & Keywords
    categories = {
        'Assets': ['asset', 'current assets', 'fixed assets', 'cash equivalents', 'inventory', 'receivables', 'property', 'investments'],
        'Liabilities': ['liability', 'debt', 'accounts payable', 'loans payable', 'taxes payable', 'borrowings', 'creditors', 'obligations'],
        'Equity': ['equity', 'shareholders equity', 'stockholders equity', 'common stock', 'retained earnings', 'net worth', 'share capital']
    }

    # ✅ Sentiment Analysis Results
    st.write("## 📝 Sentiment Analysis Results:")

    for category, keywords in categories.items():
        st.write(f"### 🔍 {category}")
        result = analyze_category(report_text, category, keywords)
        if result[0] is None:
            continue
            
        sentiment_percentages, negative_sentences = result

        # Display Sentiment Metrics
        cols = st.columns(3)
        cols[0].metric(label="✅ Positive", value=f"{sentiment_percentages['Positive']:.1f}%")
        cols[1].metric(label="⚠️ Negative", value=f"{sentiment_percentages['Negative']:.1f}%")
        cols[2].metric(label="ℹ️ Neutral", value=f"{sentiment_percentages['Neutral']:.1f}%")

        # Show Negative Sentences (if any)
        if negative_sentences:
            with st.expander("🔻 View Negative Sentences"):
                for idx, (sentence, probs) in enumerate(negative_sentences, 1):
                    st.write(f"**{idx}.** *{sentence}*")
                    st.caption(f"Probabilities → Positive: {probs[0]:.2f}, Negative: {probs[1]:.2f}, Neutral: {probs[2]:.2f}")
        else:
            st.success("✅ No negative sentences detected.")