LLM Capacity for Structured Extraction from Occupant Feedback

Companion code for the SocialSys'26 workshop paper "Towards Uncovering Indoor Satisfaction Profiles: LLM Capacity for Structured Extraction from Occupant Feedback."

This repository contains the analysis pipeline, paper source, and a synthetic demo dataset for a study that combines latent profile analysis (LPA) of occupant satisfaction ratings with locally deployed LLM extraction of structured complaint dimensions (tone, severity, attribution, work impact) from free-text feedback.

The published numerical results — eight occupant profiles, model-size thresholds for kappa ≥ 0.6 with human coders, and the classification benchmark showing text/rating orthogonality — were produced on the licensed CBE Occupant Survey database, which cannot be redistributed. The repository ships a fully synthetic dataset so that the pipeline can be exercised end-to-end on a fresh clone.

Contents

office_profiler/
├── _targets.R                  # targets pipeline definition
├── R/                          # pure functions sourced by targets
├── scripts/                    # one-shot scripts for the real-data path
├── data/
│   ├── raw/README.md           # how to obtain the real CBE database
│   └── synthetic/              # synthetic demo dataset + generator
├── paper/                      # ACM sigconf source for the workshop paper
│   ├── main-text.tex
│   ├── img/                    # figures (regenerated by tar_make)
│   └── references.bib
├── LICENSE                     # MIT (code)
├── LICENSE-paper               # CC-BY 4.0 (paper text + figures)
├── CITATION.cff                # citation metadata
├── CONTRIBUTING.md             # issue / PR guidance
├── renv.lock                   # locked package versions
└── .Rprofile                   # renv activation

Quick start (synthetic demo)

The synthetic demo runs the full pipeline on simulated data without needing access to the CBE database. Targets that call Ollama use a pre-computed embedding cache, so the demo runs CPU-only.

git clone https://github.com/IEQLab/office_profiler.git
cd office_profiler

# in R, from the repo root
renv::restore()           # install locked package versions

# regenerate the synthetic dataset (already checked in; this is optional)
source("data/synthetic/generate_synthetic.R")

targets::tar_make()       # build all figures

After tar_make() completes, the figures are written to paper/img/:

• 6_validation_kappa.png — LLM-vs-human agreement by model size
• 8_classification_comparison.png — 5-model classification benchmark
• 1_fit_comparison.png, 1_classification_quality.png, 1_split_half_ari.png — LPA diagnostics

A successful demo run takes about 3–5 minutes on a recent laptop.

Reproducing the paper results (real CBE data)

1. Request access to the CBE Occupant Survey database (see data/raw/README.md).

2. Place the export at data/raw/db_all.rds.

3. Pull the LLM and embedding models in Ollama:

   ollama pull gemma3:27b       # main extraction model (validated)
   ollama pull llama3.1:8b      # alt model variant for kappa benchmark
   ollama pull llama3.2:3b      # alt model variant
   ollama pull nomic-embed-text # embedding model

4. Run the real-data path:

   source("scripts/01-data.R")           # cleans data/raw/db_all.rds
   # the LPA-driven profile assignments must already exist at
   # data/processed/df_profiles.rds — run the LPA targets via
   # targets::tar_make(c(model_lpa, df_profiles)) and write df_profiles
   # from the model output, or supply them manually.
   source("scripts/02-llm-extraction.R") # runs Ollama, ~1 hour
   source("scripts/03-llm-validation.R") # kappa across models
   targets::tar_make()                   # rebuilds figures from real outputs

   On a workstation with a single 24 GB GPU, the gemma3:27b extraction step takes roughly 60 minutes for ~6,000 stratified responses; embedding generation is cached after the first run.

Hardware notes

• CPU-only: the synthetic demo runs in a few minutes on any modern laptop. The full real-data pipeline will work CPU-only but the gemma3:27b extraction takes many hours.
• GPU: an NVIDIA card with ≥ 16 GB VRAM (or an Apple Silicon machine with ≥ 32 GB unified memory) is recommended for the real-data path.

Citation

If you use this repository or the synthetic demo data, please cite the workshop paper (preferred) and the repository:

@inproceedings{parkinson2026towards,
  title     = {Towards Uncovering Indoor Satisfaction Profiles:
               LLM Capacity for Structured Extraction from
               Occupant Feedback},
  author    = {Parkinson, Thomas and Schiavon, Stefano and
               Zhang, Wenhao and Miller, Clayton},
  booktitle = {Proceedings of the SocialSys'26 Workshop},
  year      = {2026},
  publisher = {ACM}
}

Licence

• Code — MIT
• Paper text and figures — CC BY 4.0

Acknowledgements

The Center for the Built Environment (UC Berkeley) for access to the Occupant Survey database. ACM SocialSys'26 reviewers for their feedback.