NLP-HashLLM — Geopolitical Policy Analysis - NLP Application

Analyze public and government discourse around tariffs and trade policy, forecast economic indicators, and compare multilingual summarizers — all in one Streamlit app.

Overview

• Interactive sections for exploration and modeling:
  • Trend Analysis: Temporal sentiment by country, stance detection for statements, and India-focused topic trends.
  • Predictive Analysis: XGBoost-based forecasts for INR/USD, CNY/USD, and MXN/USD using market features and tariff-news signals.
  • Summarization: Side-by-side comparison of mT5_XLSum and Pegasus with ROUGE/BLEU evaluation and rationale highlighting.
• Built with streamlit, transformers, xgboost, scikit-learn, spacy, nltk, and sentence-transformers.

Project Structure

• code_files/app.py: Streamlit entrypoint with sidebar navigation and section routing.
• code_files/sections/:
  • trend_analysis.py: Reddit sentiment trends, XLM-R stance classifier, India NER/topic/sentiment analytics.
  • predictive_analysis.py: Exchange-rate forecasting for India, China, Mexico.
  • summarization.py: mT5_XLSum vs Pegasus summaries with ROUGE/BLEU and rationale.
• code_files/stance_xlmr_model/: Pretrained XLM-R sequence classifier (tokenizer + weights) for stance detection.
• code_files/data/: CSV datasets used across sections (news, FX, indices, labeled sentiment, topic labels, etc.).
• Notebooks and reports: NLP_Final.ipynb, NLP_Milestone.ipynb, NLP_Final_Report.pdf, NLP_Report_2 (1).pdf for methodology and experiments.

Requirements

• Python 3.10+
• Recommended: CUDA-enabled GPU for faster Transformer inference (CPU works).
• Install dependencies:

  pip install -r code_files/requirements.txt

• spaCy small English model is specified in requirements.txt. If installation fails, manually run:

  python -m spacy download en_core_web_sm

• NLTK resources (punkt, stopwords, wordnet) are downloaded automatically on first run when generating enriched NER/topic data.

Run the App

streamlit run code_files/app.py

• Open the browser link Streamlit prints.
• Use the sidebar to switch sections: Introduction, Trend Analysis, Predictive Analysis, Summarization.

Features

• Trend Analysis
  • Country-wise Reddit sentiment timeline from code_files/data/reddit_sentiment_labeled.csv.
  • Stance detection (Public vs Government) using the bundled XLM-R model in code_files/stance_xlmr_model/.
  • India news analytics: cleaned NER entities, economic term extraction, LDA topic labels, sentiment, and weekly topic trends.
• Predictive Analysis
  • INR/USD, CNY/USD, MXN/USD forecasting with engineered lags, moving averages, RSI/MACD (TA), and tariff-news signals.
  • Visual comparison of actual vs predicted with RMSE and R² metrics.
• Summarization
  • mT5_XLSum and Pegasus summaries computed on demand.
  • ROUGE and BLEU evaluation against an optional reference.
  • Rationale highlighting: shows input sentences most similar to summary content via sentence-transformers embeddings.

Key Data Files

• code_files/data/merged_exchange_tariff_data.csv: INR/USD with tariff sentiment and counts.
• code_files/data/CNY_USD_exchange.csv, code_files/data/MXN_USD_exchange.csv: FX time series.
• code_files/data/china_news_with_sentiment.csv, code_files/data/mexico_news_with_sentiment.csv: News with sentiment.
• code_files/data/India_sector_news_articles.csv: Source for India topic/NER enrichment.
• code_files/data/new_NER.csv: Generated India enriched dataset (created on first run if missing).

How It Works

• App Routing: Sidebar sets section and dispatches to analysis modules from code_files/sections/.
• Caching: Heavy models and embeddings are cached using @st.cache_resource to avoid repeated downloads and initialization.
• Modeling:
  • Stance classifier uses XLM-R loaded locally; inference runs on GPU if available, else CPU.
  • FX forecasting uses XGBRegressor with log transforms and technical indicators.
  • Summarizers use google/pegasus-xsum and csebuetnlp/mT5_multilingual_XLSum via transformers.

Tips

• If large datasets are missing, ensure code_files/data/ contains the listed CSVs.
• First run may take longer due to model and resource downloads.
• For reproducible results, pin Python and package versions and run in a fresh virtual environment.

Launch Checklist

• Dependencies installed without errors.
• code_files/stance_xlmr_model/ present and readable.
• CSVs in code_files/data/ present (or adjust paths inside sections/*).
• Start with streamlit run code_files/app.py and navigate via sidebar.