Short Drama Text Analysis: Human-AI Dialogue Behavior Mining

Project Overview

Processed internal human–AI role-play dialogues (JSON → CSV) and built conversation-level metrics to quantify practice rhythm, engagement, diversity/novelty, and session variation.

What I Did

• Parsed nested JSON dialogues into structured tabular datasets (conversation/message level).
• Engineered metrics: turns, duration, interaction density, lexical diversity, repetition and novelty proxies.
• Summarized behavioral patterns across sessions and exported clean datasets for downstream modeling.

Methodology

• Ingestion + cleaning: role separation, timestamp normalization, text normalization, invalid record removal.
• Conversation reconstruction: session grouping and sequencing; message-level to conversation-level aggregation.
• Feature engineering: interpretable indicators for engagement, rhythm, and content diversity.

Reflection

Key takeaway: conversational data is unstructured by nature, so a reliable processing pipeline is the real foundation for any dialogue analytics or evaluation.