rtc-ner-extended

rtc-ner-extended

rtc-ner-extended is a domain-specific Named Entity Recognition (NER) model developed for extracting structured information from unstructured road traffic crash (RTC) narratives. It works to identify entities not covered by the pidakwo/rtc-ner model

The model identifies RTC attributes such as rtc_time, weekday, no_vehicle, vehicle_type, persons_involved, and rtc_factors (causal factors, collision types, collision effects, physical site attributes, environmental condition, actors – humans and animals, goods types, and general vehicle category). Additionally, the model was also used as part of a data anonymization workflow in which it was used to identify entities - person, plate_no, address, and victim_organization which indicate personal details of road users involved in the RTC incident. In order to hide the personal details in the content feature, NER entities identified were replaced with their label names.

Model Description

rtc-ner-extended is an English-language spaCy NER model consisting of a tok2vec component and an NER component.

The model uses spaCy's:

• MultiHashEmbed representation for token features
• MaxoutWindowEncoder for contextual token representations
• Transition-based NER architecture for entity recognition

The model was configured and trained using spaCy 3.8.x.

Entity Labels

rtc-ner-extended recognizes the following ten domain-specific entity types:

Intended Use

rtc-ner-extended is intended primarily for research and information-extraction applications involving road traffic crash narratives.

Potential applications include:

• Extraction of structured information from RTC reports
• Construction and enrichment of RTC datasets
• Identification of crash-related factors
• Extraction of personally identifiable information (PII)
• Extraction of temporal information
• Preparation of RTC narratives for downstream machine learning
• Natural language processing of road safety reports
• Data curation and information structuring
• Support for road traffic crash analysis and research

Data Anonymization Application

rtc-ner-extended was also used as part of a privacy-preserving data processing workflow.

Four entity categories were specifically identified as potentially containing personally identifiable information (PII):

• PERSON
• PLATE_NO
• ADDRESS
• VICTIM_ORGANIZATION

These entities can be identified in an RTC narrative and subsequently replaced with their entity labels or other designated placeholders during anonymization.

Important: NER-based anonymization should not be regarded as a guarantee that all personally identifiable or sensitive information has been removed. Human review and/or additional privacy-preserving processing should be considered before publicly releasing processed text.

Model Architecture

The model uses the following spaCy pipeline:

tok2vec → ner

Model Files

The repository contains the complete trained spaCy model and its associated resources, including:

config.cfg
meta.json
tokenizer
vocab/
ner/
tok2vec/

These components should be retained together when loading the model.

Performance

The performance figures are recorded in the model's meta.json metadata.

Model Usage Test

The RTC_NER_Extended_model_test.ipynb file contains the code for testing usage of the model.

The final output contains the entities identified by the model together with their corresponding entity labels.

Limitations

rtc-ner-extended is a domain-specific research model and should not be assumed to identify every relevant entity in every RTC narrative.

Performance may be affected by:

• Differences in narrative writing styles
• Spelling and typographical errors
• Ambiguous entity boundaries
• Unusual abbreviations
• Previously unseen terminology
• Domain shifts between training and application data
• Differences in reporting conventions
• Context-dependent interpretations of entities

In particular, automated anonymization should not be considered sufficient on its own to guarantee that an RTC narrative contains no personally identifiable information.

For applications involving public release of textual data, model predictions should be complemented by appropriate validation and privacy review.

Research Context

rtc-ner-extended was developed as part of research investigating the transformation of unstructured road traffic crash narratives into structured, machine-readable information.

The model extends the information-extraction capability of the domain-specific pidakwo/rtc-ner model pipeline by identifying entities associated with crash circumstances, persons, vehicles, temporal information, and potentially privacy-sensitive information.

The extracted information can subsequently support data curation, analysis, machine learning, and road safety research.

Related Model

A separate model, pidakwo/rtc-ner, was developed for the extraction of geographic and selected incident-related entities from RTC narratives.

RTC-NER and rtc-ner-extended are separate but complementary models, and should not be treated as interchangeable.

The code for model training and evaluation as well as data extraction can be found at: https://github.com/PatUnoka/Geospatial-and-Contextual-Information-Extraction-from-Road-Traffic-Crash-Narratives.git

Citation

If you use rtc-ner in academic research, please cite the associated research publication and dataset from which the model was developed.

[1] P. O. Idakwo, O. Adekanmbi, A. Soronnadi, and A. David, “Geo-parsing and analysis of road traffic crash incidents for data-driven emergency response planning,” Heliyon, vol. 11, no. 4, p. e41067, 2025, doi: 10.1016/j.heliyon.2024.e41067.

[2] P. O. Idakwo, O. Adekanmbi, and A. David, “Nigerian Multi-modal Road Traffic Crash Data,” 2026, doi: 10.5281/ZENODO.15862127.

Disclaimer

rtc-ner-extended is provided for research purposes. Model predictions are automated outputs and should be independently validated before being used in operational, medical, emergency-response, privacy-critical, or other high-stakes applications.