Generating Récit from Sensor Data: Evaluation of a Task Model for Story Planning and Preliminary Experiments with GPS Data

Belén A. Baez Miranda, Sybille Caffiau, Catherine Garbay, François Portet

Abstract

This research explores the automated generation of "activity récits"—coherent narratives of real-world experiences—from raw sensor data such as GPS tracks and physiological metrics. The study proposes a pipeline utilizing a structured "task model" to plan narrative structures. By comparing machine-generated stories against human-written accounts, the authors demonstrate that formal task models can produce more chronologically clear and logically consistent narratives. Evaluation results show a significant reduction in timeline reconstruction errors for readers of generated texts.

Introduction

While natural language generation (NLG) has long addressed fictional storytelling, narrating real-world human activities presents unique challenges. Transforming a stream of GPS coordinates or heart rate data into a faithful, engaging story requires abstracting low-level data into high-level activities (e.g., "climbing," "resting") and organizing them into a coherent sequence. This paper addresses the "récit plan" problem: how to structure narrative sequences from messy sensor data to ensure a human reader can accurately reconstruct the original experience.

Methodology

The proposed system follows a data-to-text pipeline that integrates domain knowledge with raw data processing.

The Task Model and Pipeline

Central to the approach is a Task Model that represents domain-specific constraints (e.g., in ski touring, an ascent must precede a descent). This model serves as the story planner, ensuring logical flow.

The conceptual pipeline for generating narratives using a task model derived from human activity annotations.

The technical architecture involves several stages:

  1. GPS Segmentation: Altitude data is simplified using the Douglas-Peucker algorithm. Segments are categorized as "ascending," "descending," "moving forward," or "break" based on speed and elevation change (dénivelée).
  2. Semantic Enrichment: The system queries OpenStreetMap (OSM) to identify geographic "anchors" such as peaks or lakes, providing necessary context for the narrative.
  3. NLG Realization: Using SimpleNLG, semantic frames are mapped into grammatically correct sentences.
System architecture for the data-to-text pipeline, integrating GPS segmentation with a knowledge base and task model.

Datasets

The study utilized the Skitour.fr Corpus of human-written narratives for initial model training and the RecordMe Parallel Corpus, which contains synchronized GPS, heart rate, and breathing rate data alongside human-written accounts of the same trips.

Results and Discussion

A user study involving 18 participants compared the clarity of human-written versus machine-generated accounts. Participants were asked to reconstruct the trip timeline after reading the texts.

  • Narrative Clarity: Machine-generated stories were found to be more chronologically clear. Human writers often use ellipses or "satellite details" that, while expressive, can obscure the actual timeline.
  • Statistical Significance: Analysis of "edit distance" (the difference between a participant's reconstructed timeline and the ground truth) showed that generated texts led to significantly fewer errors ($F(1,18) = 7.583, p = 0.0131$).
  • Contextual Importance: Integrating OSM data proved vital. GPS data alone lacks the "goals" (e.g., reaching a peak) that provide narrative satisfaction.

Conclusion and Future Work

The research validates the use of formal task models for ensuring narrative coherence in data-to-text systems. Future work aims to scale the parallel corpus and incorporate physiological data—such as heart rate—to detect emotional or physical states like "exhaustion," adding depth to the generated récits. Additionally, the authors plan to improve linguistic flow by adding discourse connectors to better express causality and handle multi-agent narratives.

Summary updated

BibTeX
@inproceedings{miranda2015generating,
  author = {Belén A. Baez Miranda and Sybille Caffiau and Catherine Garbay and François Portet},
  title = {{Generating Récit from Sensor Data: Evaluation of a Task Model for Story Planning and Preliminary Experiments with GPS Data}},
  booktitle = {ENLG},
  year = {2015},
  doi = {10.18653/v1/w15-4714},
}
Figure