About NDL Face
NDL Face is a research platform for exploring descriptions of facial expressions, emotions, and appearance extracted at scale — using large language models — from the OCR full text of literary works in the NDL Digital Collections.
Collecting emotional expressions from literary texts has traditionally depended on close reading by researchers, which limits coverage and speed as the corpus grows. This project addresses that with a pipeline that runs multiple LLM agents in parallel over hundreds of pages of classical and modern Japanese text.
The project is affiliated with the JSPS KAKENHI project "A History of the Correlation between Appearance Description and Inner Evaluation" (Grant Number 25K21831) at the University of Tokyo.
Features
- LLM multi-agent extraction — structured records of expression (surface) and context that morphological analysis alone cannot capture.
- Multi-axis categorisation — 11 emotion categories, body parts (face, eyes, mouth, cheeks, …), valence (positive/negative), and intensity.
- Faceted search — by emotion, body part, valence, intensity, character gender, genre, author, and publication year, plus keyword search.
- Linked to the original images — each expression links to the corresponding frame in the NDL Digital Collections, with the matched text region overlaid in the IIIF viewer.
Pipeline
- Fetch OCR full text from the NDL Digital Collections
- Extract expressions (surface + context) with parallel LLM agents
- Normalise emotion / body part / valence / intensity to a controlled vocabulary
- Enrich bibliographic metadata (author, year, NDC) from NDL IIIF manifests
- Match OCR coordinates to generate bounding-box regions on the source images
- Index everything into Elasticsearch for cross-corpus search
Data & licensing
- Source images and full text: NDL Digital Collections (images are served via the NDL IIIF Image API and are not hosted by this site).
- The extracted data is published for research purposes. Extraction is LLM-based and may contain errors.