#AI-Powered Sentence Completion for Video Game Narratives

This project represents my exploration into the fascinating intersection of artificial intelligence and interactive storytelling. As video games continue to evolve as a narrative medium, I wanted to investigate how modern NLP techniques could understand, predict, and generate contextually relevant text within gaming environments.

Video games aren't just entertainment—they're immersive narrative experiences that shape player emotions and engagement. This project explores how AI can enhance and co-create these narratives by building an intelligent sentence completion system trained on diverse video game texts.

• Hand-curated dataset of video game narratives with emotional context annotations
• Diverse text sources including game scripts, lore entries, and narrative content
• Emotional labeling system for mood-aware text generation
• Structured data pipeline for scalable text processing

• Text Preprocessing: Advanced tokenization, normalization, and feature extraction
• Sequence Generation: Sliding window approach creating 5-word input sequences
• Feature Engineering: TF-IDF vectorization for meaningful text representation
• Model Training: Multinomial Naive Bayes with probabilistic sampling for diverse outputs

• Contextual Prediction: Next-word prediction with 5-word context windows
• Creative Generation: Auto-generates plausible game-style sentences
• Emotion Integration: Emotion-aware text generation for narrative tone control
• Planned Interface: A Streamlit demo was scoped for real-time sentence completion but not built — the model currently runs from the notebook only

Core Technologies: Python, Scikit-learn, NLTK, Pandas, NumPy Visualization: Matplotlib, Seaborn Development Environment: Jupyter Notebook Deployment: Notebook-only — Streamlit was scoped as future work, not shipped NLP Techniques: TF-IDF, Naive Bayes, Text Generation

The system demonstrates sophisticated understanding of gaming narrative patterns:

• Contextual Accuracy: Successfully predicts appropriate next words based on gaming context
• Creative Output: Generates grammatically correct, thematically consistent sentences
• Emotional Awareness: Incorporates emotional context for more nuanced text generation

Input: "the hero must" → Output: "the hero must find the lost sword"
Input: "after the battle" → Output: "after the battle the village was silent"
Input: "choose your weapon" → Output: "choose your weapon wisely before the fight"

This project showcases several advanced AI engineering concepts:

Unlike using pre-existing datasets, I engineered a specialized corpus with emotional annotations, demonstrating data science expertise and domain knowledge.

From raw text collection through model training and generation, this project represents complete ownership of the ML lifecycle.

Applying NLP to creative content generation shows understanding of AI's potential beyond traditional analytics.

Pipeline (single-notebook, no live service):

Raw game-text corpus
  → Punctuation stripping + tokenization (NLTK word_tokenize)
  → Sliding-window dataset construction (5-word context → 6th word target)
  → TF-IDF vectorization of the context strings
  → Multinomial Naive Bayes classifier (context vector → next-word class)
  → Top-5 probability sampling (not greedy) for varied completions
  → Autoregressive generation: predict → append → slide window → repeat

Key modeling decisions:

• Framed next-word prediction as multi-class classification over the vocabulary, not a language-model objective — TF-IDF + Naive Bayes is a deliberately lightweight baseline that demonstrates full pipeline ownership (curation → features → model → generation) without needing GPU training infrastructure.
• Top-5 sampling instead of greedy argmax — always taking the single most probable word produces repetitive output ("the and the and..."); sampling among the top 5 trades a little precision for noticeably more human-like variety.
• Custom emotion-annotated dataset — training text was hand-curated from game scripts and lore with emotional labels, laying groundwork for mood-aware generation. The labels exist in the data but aren't yet a model input.
• Streamlit interface was scoped, not built — the README described a real-time demo, but the model runs from the notebook only. Listed here as future work rather than a shipped feature.

Documented next steps: swap TF-IDF/Naive Bayes for transformer embeddings for real contextual understanding; extend from single-sentence to paragraph/dialogue generation; wire the emotion labels in as an actual conditioning signal; build the Streamlit interface.

Advanced Architectures: Integration of transformer models (BERT, GPT) for improved context understanding Extended Generation: Expansion to paragraph and dialogue-level narrative creation Emotion-Driven Control: Enhanced emotional modeling for narrative tone and pacing control Production Deployment: Full web application with user feedback integration

✅ Complete NLP Pipeline: End-to-end development from data collection to text generation ✅ Custom Dataset Engineering: Specialized corpus creation with emotional annotations ✅ Modular Architecture: Scalable, well-documented codebase for future enhancement ✅ Creative AI Application: Demonstrates AI's potential in content generation and storytelling

This project deepened my expertise in:

• Advanced NLP preprocessing and feature engineering
• Probabilistic modeling and text generation techniques
• Custom dataset creation and annotation
• Interactive ML application development
• Creative applications of traditional ML algorithms

The combination of technical rigor and creative application makes this project a compelling demonstration of AI engineering skills applied to real-world narrative challenges.

This project represents my commitment to exploring AI's creative potential while maintaining technical excellence and practical applicability.