Planning Presentation v1

1. Project Scope & Objectives

• Build a research-focused argument analysis system for political opinions.
• Users post face-camera opinion videos on trending political topics.
• AI classifies each video reply as FOR, AGAINST, or UNRESOLVED.
• Every AI classification is linked to an exact transcript excerpt and timestamp.
• Evaluate the system against baselines using F1, Precision, Recall, and Fleiss' Kappa.

2. Proposed Functions & Intended Users

• Content Creators: Post opinion videos on trending political topics.
• Viewers: Navigate the visual argument graph to understand all sides of a debate.
• Community Moderators: Vote on UNRESOLVED classifications and report misleading content.
• Researchers: Access structured argument data for analysis.

3. System Features

• AI Stance Classification: Automatic FOR / AGAINST / UNRESOLVED labeling via LLM.
• Timestamp Traceability: Every classification cites the exact video time range.
• Multi-Dimensional Scoring: Fact Score, Truth Rating, Emotional Appeal, Comedy/Satire.
• Community Voting: Unresolved videos go to a public voting queue.
• Reporting & AI Moderation: Flag factless/misleading videos; AI re-verifies on threshold.
• Interest-Based Recommendations: AI-generated topic tags and hashtags.

4. Architecture & Tech Stack

• Frontend (Web): React.js with D3.js for graph visualization.
• Frontend (App): Flutter (Dart).
• Backend: Spring Boot (Java) for the main FORREST platform.
• Team Website Backend: Node.js + Express.
• Database: Neo4j (Graph DB for argument trees) + PostgreSQL (user data).
• Speech-to-Text: OpenAI Whisper (open-source).
• Stance Classification: Google Gemini API / OpenAI GPT-4o API.
• Fact-Checking: RAG (Retrieval-Augmented Generation) via LangChain.
• Deployment: Docker containers on a cloud VPS.

5. Evaluation & Baselines

• Baseline 1 (Zero-Shot): Raw transcript → LLM → FOR/AGAINST classification.
• Baseline 2 (Summary-Only): Transcript summary → LLM → classification.
• FORREST Pipeline: Parent context + transcript + structured extraction.
• Metrics: Precision, Recall, F1-Score, Stance Accuracy, Timestamp Overlap (IoU).
• Human Agreement: Fleiss' Kappa across 4 annotators on 20-video pilot dataset.

6. Technical Risks & Challenges

• Sarcasm & Ambiguity: AI may misclassify sarcastic videos. Mitigated by UNRESOLVED + community voting.
• API Costs: LLM API calls have per-token costs. Mitigated by using free-tier quotas and caching.
• Low Annotator Agreement: If humans disagree on stances, AI evaluation is unreliable. Mitigated by refining annotation guidelines.
• Video Quality: Poor audio quality degrades transcription. Mitigated by requiring face-camera videos with clear speech.

7. Gantt Chart

The detailed Gantt chart with tasks, dependencies, milestones, and deadlines will be embedded here.

(To be updated with the interactive Gantt chart before the planning presentation.)