BRIEFING.AI

Extractive text summarization engine tailored for long, complex terms-of-service and legal documents. Built using classical Natural Language Processing (NLP), Machine Learning graph centrality, and a hyper-minimalist, brutalist Streamlit interface.

📊 Project Presentation: Download PDF Presentation (If the GitHub preview fails, download/view the raw PDF file directly).

---

FEATURES

• Pure Classical NLP & ML: Sentence tokenization and lemmatization via spaCy, TF-IDF vectorization, Cosine Similarity matrices, and PageRank scoring (via NetworkX).
• No LLM/Transformer API Calls: Zero external model dependencies or api-keys; runs 100% locally and securely.
• Custom Legal Preprocessing: Specialized stop word filters for legalese (e.g., whereas, hereby, heretofore, covenants).
• Brutalist Monochrome Aesthetic: Stark pitch black interface with silver/white monospace text, professional metrics, and zero visual clutter.
• Flexible Controls: Select summary lengths by absolute sentence count or target compression ratio.
• Detailed Evaluation Framework: Built-in benchmark suite analyzing execution time, sentence compression ratios, PageRank score distributions, and sentence connection topology graph structures.

---

QUICK SETUP

1. Clone the Repository

git clone https://github.com/Nirali4/Briefing.AI.git
cd Briefing.AI

2. Create and Activate a Virtual Environment

On Windows:

python -m venv .venv
.venv\Scripts\activate

On macOS/Linux:

python3 -m venv .venv
source .venv/bin/activate

3. Install Dependencies

pip install -r requirements.txt
# Additional libraries for evaluation, plotting, and reading files:
pip install matplotlib pandas seaborn pypdf python-docx ipykernel

4. Download the spaCy Language Pipeline

python -m spacy download en_core_web_sm

---

HOW TO USE

1. Launch the Streamlit Server

streamlit run app.py

Or, if .venv is not activated

.venv\Scripts\streamlit.exe run app.py

This will start the local server and open the web app in your browser (usually at http://localhost:8501).

2. Run Summarization

• Input Text: Either upload a plain text, PDF, or DOCX legal document using the uploader, or copy-paste the contract/TOS text directly into the primary input area.
• Configure Length: Use the configuration controls to choose between Sentence Count or Compression Ratio and adjust the slider to your desired summary length.
• Execute: Click RUN SUMMARIZATION ENGINE to compute sentence centrality. The top N central sentences will be extracted and presented in their original chronological sequence alongside compression metrics.

---

PERFORMANCE & EVALUATION

A comprehensive quantitative evaluation suite has been created to analyze the summarizer's accuracy, execution efficiency, and centrality distribution across different legal payloads:

• Jupyter Notebook: Evaluation_and_Results.ipynb details dataset profile metrics, legal preprocessing challenges, execution benchmarks, and network graphs.
• Benchmark Data: performance_metrics.csv contains character, word, sentence profiles and processing times for the files in the test/ folder.
• Generated Visualizations (located in the assets/ directory):
  • performance_plot.png: Line plot representing document scale (sentence count) vs. preprocessing and algorithmic execution times.
  • compression_plot.png: Bar chart comparing original word counts to summarized word counts.
  • pagerank_distribution.png: Score distribution indicating selection boundary clarity.
  • sentence_network.png: Topological graph displaying connectivity between sentences, highlighting top central nodes in green.

To re-run the benchmark suite and generate the notebook outputs:

python scratch/run_evaluation.py

---

SYSTEM ARCHITECTURE

[Input Document] 
      │
      ▼
[spaCy Pipeline] ──► Sentence Tokenization & Custom Stop Words Filtering
      │
      ▼
[scikit-learn]   ──► TF-IDF Sparse Matrix Construction
      │
      ▼
[Cosine Sim]     ──► Matrix Pairwise Cosine Similarity Computation
      │
      ▼
[NetworkX]       ──► PageRank centrality computed over Sentence Similarity Graph
      │
      ▼
[Sort & Output]  ──► Extraction of Top N nodes mapped back chronologically