Smalltalk-KG: A Knowledge Graph Construction Framework for Dialog Personalization in Human-Robot Interaction

Overview

Smalltalk-KG is a local knowledge graph framework for long-term dialog personalization in human-robot interaction. It models conversational memory as time-stamped episodic knowledge and stable facts, enabling low-latency, personalized, and traceable responses without relying on long prompt histories or cloud-based storage.

The system is designed for socially assistive robotics and gerontological HRI, where consistent long-term memory, explainability, and local-only data processing are especially important.

If you decide to use our work, please cite it as follows:

@inproceedings{wils2026smalltalkkg,
  title={Smalltalk-KG: A Knowledge Graph Construction Framework for Dialog Personalization in Human-Robot Interaction},
  author={Wils, Niklas and Peral, Sergio and Blum, Kai Michael and Gerndt, Reinhard and Doernbach, Tobias},
  booktitle={International Conference on Advanced Robotics and its Social Impact},
  year={2026},
  note={accepted}
}

Pipeline Runner

Purpose

The pipeline performs two stages:

1. Simulation of conversational episodes.
2. Visualization of the resulting knowledge graph.

Requirements

• Install dependencies from requirements.txt.

Quick Start

Default simulation + visualization:

python scripts/run_full_pipeline.py

Common Experimental Controls

Configure the simulation scale:

python scripts/run_full_pipeline.py \
  --persons 10 \
  --weeks 8 \
  --sessions_per_week 3 \
  --turns_per_session 12 \
  --seed 42

Redirect outputs:

python scripts/run_full_pipeline.py \
  --out_root runs \
  --viz_output_dir results/visualizations

Constrain visualization size:

python scripts/run_full_pipeline.py \
  --max_episodes 800 \
  --max_episodes_per_topic 60

Visualization only (latest run):

python scripts/run_full_pipeline.py --skip-simulation

Visualization only (specific run):

python scripts/run_full_pipeline.py --skip-simulation --run_id 20260204_203710_seed42

Outputs

• Simulation output: runs/<run_id>/kg_store.nq
• Visualization output: results/visualizations/kg_store_<run_id>_galaxy.png

Example Visualization

The figure below shows an example knowledge graph visualization produced by the pipeline (galaxy layout).

Notes for Reproducibility

• --skip-simulation without --run_id selects the most recent run in --out_root.
• When --out_root changes, visualization inputs are read from the same location.
• Record the full command and --seed value in experiment logs.