Ph.D. Candidate @ KAIST Graduate School of Culture Technology
Computational Humanist.

Building pipelines that turn unstructured text into measurable structure.

I design end-to-end ML pipelines for unstructured text data — from large-scale web crawling and OCR preprocessing to embedding-based analysis and LLM-powered applications.

My research sits at the intersection of Natural Language Processing and Computational Humanities, with a focus on:

• Semantic trajectory modeling in high-dimensional embedding space
• LLM-based surprisal estimation for narrative structure analysis
• RAG pipeline design for domain-specific text retrieval
• Multimodal (text + image) association frameworks

Languages Python R SQL Cypher

ML / NLP PyTorch HuggingFace Transformers Sentence-Transformers scikit-learn XGBoost LightGBM

LLMs & Embeddings Solar-10.7B (4-bit quantized) Ko-SRoBERTa Whisper Tesseract OCR

Data & Infra FAISS pandas NumPy Selenium MySQL Neo4j Conda/pip

Tools Jupyter Git Linux (Ubuntu) NVIDIA GPU (RTX A6000)

Computational analysis of narrative structure in 2,636 Korean flash fiction stories using sentence-level signal trajectories.

• Ran Solar-10.7B (4-bit NF4 quantized) on GPU server for sentence-level surprisal estimation across full corpus
• Computed coherence (Ko-SRoBERTa cosine similarity) and semantic shift (cumulative context deviation)
• Identified 5 structural archetypes via k-means clustering on normalized 50-point trajectories
• Validated "fracture and leap" cycle: surprisal peaks → coherence drop + semantic shift spike (Wilcoxon, Bonferroni-corrected)

Solar-10.7B Ko-SRoBERTa k-means KDE Kruskal-Wallis bitsandbytes

Geometric differentiation of two visually indistinguishable short-form prose genres via semantic trajectory analysis.

• Constructed a 5D coordinate system (Displacement, Surprisal Volatility, Global Resonance, Curvature, Spectral Centroid) from sentence embeddings
• Applied PERMANOVA (pseudo-F = 13.66, p < 0.001) and Levene's test to confirm location-driven separation
• Validated coordinate independence: VIF < 1.2 across all 5 dimensions
• Ablation study (Random Forest probe) confirmed Surprisal Volatility as the strongest discriminating feature

Ko-SRoBERTa PERMANOVA OLS regression Random Forest SHAP-style ablation

Large-scale word–color association via automated web image retrieval and CIELch color space analysis.

• Built a fully reproducible pipeline: Google Image crawling (Selenium) → pixel-level color extraction → 10-dimensional color vector per word
• Applied to top 100 global brands (cosine similarity between logos and web images) and 20th-century Anglo-American poetry (Amy Lowell vs. Georgia Douglas Johnson)
• Language-agnostic framework; open-source module included

Python Selenium scikit-image Bootstrap resampling CIELch

Part of KOCCA-funded broadcasting monitoring R&D (2023–2025)

Robust speech transcription pipeline with vocal separation for noisy/music-heavy audio.

• Demucs (HTDemucs) for source separation with automatic fallback logic (RMS threshold, duration check)
• Whisper large-v2 for timestamped transcription
• Handles .mp3 / .mp4; designed for large-scale broadcast monitoring

Demucs Whisper FFmpeg PyTorch

각 논문의 최신 상태는 아래 표가 유일한 기준입니다.

• Ph.D. Cultural Technology, KAIST (2022–2027 expected)
• M.Eng. Cultural Technology, KAIST (2020–2022)
• B.Eng. Industrial & System Engineering + Culture & Arts Software + Creative Writing, Dongguk Univ. (2015–2020)

📧 wangsp0317@kaist.ac.kr · 🌐 spwang317.github.io