Pith. sign in

REVIEW 6 major objections 6 minor 30 references

Benchmarked against official state crash coding, six frontier LLMs failed to beat a keyword-rule baseline or the always-majority baseline on raw agreement, and the paper concludes that claims these models add value over conventional text pr

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-03 14:18 UTC pith:PAB2CD75

load-bearing objection Careful negative-result benchmark on LLM crash coding, but unvalidated record linkage and missing reproducibility keep the central claim conditional. the 6 major comments →

arxiv 2607.29064 v1 pith:PAB2CD75 submitted 2026-07-31 cs.LG cs.AI

Benchmarking Frontier Large Language Models Against Official Crash Database Coding Using Police Crash Narratives

Sudhir Bharati , Rajendra K C Khatri , Sudip Bharati This is my paper
classification cs.LG cs.AI
keywords LLM benchmarkcrash narrative codingcrash databaseszero-shot extractionCohen's kapparecord linkageabstentionroad safety
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper tries to establish whether frontier LLMs can reproduce official crash-database coding from police crash narratives, using 4,194 linked fatal crashes from Arkansas (2015–2025) and six attributes coded zero-shot. The results undercut the case for deployment: the best LLM (GPT-5.5 High) reached 38.5% raw agreement and kappa 0.275, while a simple keyword-rule system reached 57.1% agreement and kappa 0.264, and the always-majority baseline reached 76.5% raw agreement. Agreement varied more across attributes (within-model spread up to 0.971) than across models, with near-zero agreement on light condition, roadway surface, and work-zone relation driven mostly by models returning an abstention code (999) that was scored as disagreement. The paper's contribution is a benchmark protocol—attribute-specific evaluation, transparent baselines, explicit abstention measures—plus a warning that any claim of LLM value over conventional text processing for this task is unsupported by the evidence.

Core claim

On its own terms, the central finding is that LLM narrative coding agrees with the official crash database too poorly for deployment: best model (GPT-5.5 High) raw agreement 0.385, kappa 0.275; a pre-specified keyword-rule baseline reached agreement 0.571, kappa 0.264; always-majority reached raw agreement 0.765. The paper concludes that any claim these models add value over conventional text processing is unsupported. Attribute agreement ranged from 0.975 (non-motorist relation) and 0.788 (crash manner) down to 0.005–0.014 (light condition, roadway surface, work-zone relation), where all models abstained (code 999) on nearly every record, so unanimity was largely shared abstention, not shar

What carries the argument

Three mechanisms carry the argument. (1) An identical zero-shot prompt with an explicit 999-abstention code, run over annual narrative files with cross-conversation memory disabled—this defines the task and produces the abstention behavior that drives low agreement on environmental attributes. (2) A metric set built around the raw-agreement/kappa inversion: raw agreement, macro-F1, Cohen's kappa (chance-corrected agreement), coverage (share of specific codes), and selective agreement (agreement on committed outputs). (3) Transparent baselines—always-majority, always-Unknown, and ordered keyword regular expressions with majority-class defaults (999 for unknown light/surface), fixed before any

Load-bearing premise

The load-bearing premise is that the record linkage is accurate—narratives paired to structured records on exact date, reported time within ±5 minutes, and roadway-name text, with no manually verified sample to estimate linkage accuracy—and that one uncontrolled run per model on consumer web interfaces yields comparable, repeatable outputs; the paper states both caveats. If linkage errors are non-random or the single runs are unrepresentative, every agreement estimate, kappa,

What would settle it

Re-run the protocol on 500 randomly drawn crashes with a manually verified linkage and a human-adjudicated reference: two blinded coders assign the six attributes from the narratives with a third coder resolving disagreements, and LLM outputs are generated with temperature 0 via documented provider APIs. If human-machine kappa on light condition, roadway surface, or work-zone relation exceeds 0.5 on narratives that explicitly state the condition, or if GPT-5.5 High beats the keyword baseline by more than 0.05 kappa at matched coverage, the paper's conclusion that LLMs add no value over convent

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Agencies should not deploy these models to populate structured crash fields: coverage topped out at 43.2%, and light condition, roadway surface, and work-zone relation showed near-zero agreement because models abstained on almost every record.
  • Reporting only raw agreement or only kappa would mislead—the two rank these systems in opposite orders; both metrics plus the always-majority baseline are needed for a fair comparison.
  • Non-motorist relation and crash manner are the only fields with LLM agreement high enough (up to 0.975 and 0.788) to justify a limited, human-adjudicated pilot in real agency workflow.
  • A conventional keyword system (kappa 0.264) must be the bar: any LLM deployment claim needs to show gains in accuracy, coverage, maintainability, or staff effort against such transparent baselines, not against raw model scores.
  • Model rankings are attribute-specific; a model chosen on overall performance will not be the best for every field, so procurement or pipeline decisions need per-attribute benchmarks.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The near-zero floor on light/surface/work-zone may be a property of the prompt's abstention threshold rather than the narratives: a direct extension would instruct models to infer conditions from implicit cues (rain implies wet surface) and compare selective agreement at matched coverage.
  • The paper's own proposed next step is the falsifier: a stratified human-adjudicated sample (roughly 400–600 crashes) with blinded reviewers would convert this agreement benchmark into an accuracy benchmark and reveal whether abstention is warranted.
  • Because the narrative and structured codes come from the same police report, the reported agreement is probably an upper bound on agreement against an independent ground truth; external validation would likely lower, not raise, the LLM numbers.
  • The unvalidated record linkage is the quiet risk: exact-date/±5-minute/roadway-name pairing has no manual-check estimate, so a non-random linkage error would bias every table; a manual re-link of a few hundred pairs, reported by attribute, would bound that bias.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

6 major / 6 minor

Summary. The paper benchmarks six frontier LLMs on the task of coding six crash attributes (crash manner, light condition, intersection type, work-zone relation, roadway surface condition, non-motorist relation) from 4,194 Arkansas fatal-crash narratives linked to official structured records. Using a shared zero-shot prompt, the authors compare model outputs with official coding via agreement, macro-F1, Cohen's kappa, coverage, and selective agreement, and against always-majority, always-Unknown, and keyword-rule baselines. They report that GPT-5.5 High achieves the highest raw agreement among LLMs (0.385), but always-majority achieves 0.765 and the keyword baseline achieves comparable kappa (0.264 vs. 0.275) and higher macro-F1 (0.226 vs. 0.211). Agreement varies more across attributes than across models, with near-zero agreement on light condition, roadway surface, and work-zone relation, driven largely by shared abstention (999). The authors conclude that the results do not support claims that LLMs add value over conventional text processing for this task and recommend attribute-specific, human-adjudicated evaluation before deployment.

Significance. If the findings hold, the paper makes a useful contribution to transportation safety informatics by providing a transparent, attribute-level agreement benchmark rather than a single accuracy claim. Its strengths are the explicit scoping to agreement (not accuracy), pre-specified keyword and majority baselines, the exact-time sensitivity analysis, bootstrap clustered at the crash level, and the careful discussion of abstention and selective agreement. The paper also gives concrete practical guidance: agencies should benchmark against transparent baselines and evaluate attributes separately. The main result — that simple baselines are competitive with frontier LLMs — is important and should be published if the linkage uncertainty and run-to-run variability are adequately addressed.

major comments (6)
  1. [Methodology, Data] The central 'no added value' claim rests on agreement estimates whose validity depends on record linkage accuracy. The authors state that the files shared no common crash identifier and that no independent manually verified validation sample was created, so a separate estimate of linkage accuracy could not be calculated. The exact-time sensitivity analysis (agreement changed by <0.4 pp, ranking unchanged) mitigates time-error bias, but it does not address roadway-name ambiguity or differential linkage errors among the 1,344 matches differing by 1–5 minutes. If mismatches are correlated with, for example, work-zone status or rural roadway naming, the near-zero agreement on work-zone relation and the attribute comparisons could be distorted. A manual validation sample stratified by time difference and roadway-match type, or a linkage-error sensitivity analysis, is needed to support the cen
  2. [AI-Based Extraction Protocol] The model-comparison component is load-bearing for the ranking claims (e.g., pairwise differences of 8.7–10.5 percentage points), yet each annual file was processed only once through paid consumer web interfaces with default sampling settings; temperature, top-p, seed, and provider-side configuration were neither controlled nor recorded, and run-to-run variability was not measured. The manuscript acknowledges this limitation, but the stated finding that GPT-5.5 models 'generally achieved higher agreement' than Claude models by 8.7–10.5 pp could still be within run-to-run variability. Repeated runs on a subsample, or at least a sensitivity analysis treating web-interface output as a random draw, would make the ranking claims more secure. Absent this, the model-ranking conclusions should be framed as exploratory.
  3. [Discussion / Implications] The sentence 'Any claim that these models add value over conventional text processing for this task is not supported by these results' is appropriately hedged by the preceding limitations, but it is stronger than the evidence warrants given the linkage and single-run issues. The keyword baseline itself has important design choices (ordered rules, default categories, 999 defaults) that were fixed a priori, which is a strength; however, the comparison is relative to one particular crude rule set, not conventional text processing generally. The manuscript would be more precise if the conclusion were explicitly conditional on the linkage assumptions and on the specific baselines and prompt used, rather than phrased as a general statement about conventional text processing.
  4. [Performance Evaluation / Results] The claim that agreement varies more across attributes than across models is based on a comparison of within-model ranges (0.577–0.971) versus within-attribute ranges (0.001–0.438). This is descriptively true in Table 2, but the near-zero agreement for light condition, roadway surface, and work-zone relation is mechanically driven by shared abstention (models returned 999 on nearly all records). The authors acknowledge this in the text, but the range comparison itself does not establish that attribute-specific benchmark design is the primary takeaway; it reflects a nearly degenerate coding behavior on those attributes. The GEE interaction test shows statistical significance, but the practical interpretation would be clearer if the range comparison were accompanied by an analysis that separates abstention-driven disagreement from specific-code disagreement.
  5. [Limitations] The limitations section is unusually candid and should be kept. However, it states that matched and unmatched records were compared only on narrative length. Given that about 25% of narratives were unmatched, the representativeness of the analytic sample is not established. If unmatched records differ systematically in crash complexity or attribute mix, the reported agreement levels may not generalize to the full fatal-crash population. The authors should either provide more detailed comparisons (e.g., available structured fields for unmatched records, if any) or explicitly state that the benchmark applies only to the linkageable subset.
  6. [Statistical Analysis] The repeated-measures ANOVA and GEE are reported with p < 0.001 and large F/chi-square values. Given 4,194 paired crashes and 150,984 observations, these significance levels are unsurprising. The paper appropriately reports absolute percentage-point differences alongside tests, which is good. However, the composite score used for the repeated-measures ANOVA takes only seven distinct values and averages six attributes with very different class structures; the confirmatory GEE on the binary outcome is the more appropriate model. The authors should state explicitly in the Results section that the composite-score ANOVA is descriptive and that the GEE is the primary inferential analysis, to avoid readers overinterpreting the ANOVA's partial eta-squared.
minor comments (6)
  1. [Figure 1] Typo: 'Methodolgical Workflow' should be 'Methodological Workflow'.
  2. [Literature Review] Reference [13] is cited twice in the same sentence: '...critical pre-crash event [13], [13].' One citation should be removed.
  3. [Results, Exploratory Patterns] Grammar: 'The association was modest may reflect unmeasured differences...' Missing 'and' or dash; likely intended 'The association was modest and may reflect...'.
  4. [Table 1 caption] The caption says 'averaged over the six attributes' but coverage is pooled across 25,164 attribute values rather than averaged per the text. Please clarify in the table or caption whether coverage is pooled or averaged; the text says pooling is used.
  5. [Abstract / Conclusion] The abstract says 'yielding 4,194 matched crashes' and the full text reports 4,230 linked pairs, of which 4,194 had fully numeric codes. This is fine, but the abstract could briefly mention that the analysis sample is the 4,194 complete-case subset.
  6. [Prompt details] The paper states that the prompt directed models to return 999 when an attribute could not be determined, but the exact wording of the prompt is not provided. Given the centrality of the 999/abstention behavior to the results, including the full prompt in an appendix would improve reproducibility.

Circularity Check

0 steps flagged

No significant circularity: the paper is an empirical benchmark with independently specified baselines and no derived prediction whose output is equivalent to its input.

full rationale

This paper is a measurement study rather than a derivation: it links police crash narratives to structured records and compares LLM outputs against official codes and against transparent baselines. There is no fitted parameter later relabeled as a prediction. The keyword baseline was specified before examining LLM results and was not revised based on observed performance, as the paper states: "The keyword rules and default categories were specified before examining the LLM results and were not revised based on the observed comparative performance." The always-majority baseline's high raw agreement is a by-construction artifact of class imbalance and the decision to score abstentions (999) as disagreement; the paper discloses exactly this rather than presenting it as an empirical discovery: "This is a direct consequence of class imbalance combined with the models' abstention behavior." No load-bearing claim is justified by self-citation; the one self-citation (reference [6], a review of protected bike lanes) is incidental background for why non-motorist and intersection attributes matter. The acknowledged absence of an independently validated reference and the unvalidated record linkage are validity threats to the benchmark's interpretation, but they are not circularity because the agreement calculations do not assume the conclusion they support. The paper does not rename a known result or smuggle in an ansatz via citation. Therefore no circularity is present.

Axiom & Free-Parameter Ledger

3 free parameters · 3 axioms · 0 invented entities

The paper's claims are empirical measurements, not derivations, so the ledger records procedural and interpretive premises rather than fitted constants. No free parameter is secretly fit to maximize performance: the keyword rules are stated to be a priori, the linkage tolerance is a design choice, and the abstention-scoring rule is disclosed and supplemented by selective agreement. The main premises are that the official database is a usable comparison standard (explicitly flagged as not independently validated), that narrative and structured fields can be compared despite sharing an origin, and that consumer-web-interface outputs with uncontrolled settings can be compared across models.

free parameters (3)
  • Record-linkage time tolerance (±5 minutes) = ±5 min
    Hand-chosen design tolerance for linking records that share no common identifier; affects the composition of the 4,194-crash sample and hence all agreement estimates. Not fitted to data.
  • Abstention scoring rule (999 counted as disagreement when official code is specific) = 999 = disagreement vs. specific official code
    Hand-chosen scoring convention that makes always-majority (0.765) beat every LLM on raw agreement; disclosed and supplemented by selective-agreement analysis, but it shapes the headline near-zero agreement on light, surface, and work-zone attributes.
  • Keyword-baseline rules and default categories = Six attribute-specific regex rule sets with defaults (e.g., default Single Vehicle, default No, default 999)
    Authored a priori 'before examining the LLM results' as a deliberately crude transparent comparator; not fitted to maximize performance, but the default choices (e.g., always-Single-Vehicle for crash manner) drive much of its raw-agreement advantage.
axioms (3)
  • domain assumption The official structured database is a usable comparison standard for agreement.
    Methodology states it is 'the comparison source/standard for these comparisons but is not an independently validated one'; disagreement cannot be attributed to model vs. database vs. linkage error. Load-bearing for interpreting every agreement number.
  • domain assumption Narrative and structured fields, though from the same police report, can be meaningfully compared.
    Methodology notes the two files 'are not independent observations because both originate from the same police report,' which may inflate agreement; the paper treats this as a conservative caveat but relies on the comparison throughout.
  • domain assumption Six LLM outputs are comparable despite uncontrolled sampling settings and consumer-web-interface defaults.
    AI-Based Extraction Protocol: temperature, top-p, seed, and provider-side configuration 'were neither controlled nor recorded'; single run per model. The model-ranking claims depend on this premise, which the paper acknowledges limits replication but still reports rankings.

pith-pipeline@v1.3.0-daily-deepseek · 14331 in / 12135 out tokens · 119285 ms · 2026-08-03T14:18:25.273343+00:00 · methodology

0 comments
read the original abstract

Police crash narratives contain information that may supplement structured crash databases, but manual review is labor-intensive and it remains unclear how well large language models (LLMs) reproduce official crash coding. This study benchmarked six frontier LLMs by comparing narrative-derived crash attribute codes with corresponding fields in the Arkansas fatal-crash database. The analysis linked 5,587 fatal-crash narratives with 5,889 structured crash records from Arkansas (2015-2025), yielding 4,194 matched crashes. Six LLMs were evaluated using an identical zero-shot prompt to code crash manner, non-motorist relation, intersection type, work-zone relation, roadway surface condition, and light condition. Performance was evaluated using agreement, macro-averaged F1 score, Cohen's kappa, coverage, selective agreement, and comparisons with always-majority, always-Unknown, and keyword-rule baselines. Repeated-measures analyses and a generalized estimating equations model assessed differences among models and attributes. GPT-5.5 High achieved the highest agreement among the evaluated LLMs, but the always-majority baseline produced higher raw agreement and the keyword-rule baseline achieved macro-averaged F1 score and Cohen's kappa comparable to the best-performing LLM. Agreement was highest for non-motorist relation and crash manner and lowest for light condition, roadway surface condition, and work-zone relation. Differences across crash attributes exceeded differences across models. These results provide a benchmark for evaluating LLM-based crash coding and show that deployment should be evaluated on an attribute-specific basis using transparent baselines and human review.

Figures

Figures reproduced from arXiv: 2607.29064 by Rajendra K C Khatri, Sudhir Bharati, Sudip Bharati.

Figure 2
Figure 2. Figure 2: Macro-F1 by model and attribute Model and Attribute Differences in Agreement The statistical analysis examined whether agreement with the official crash database differed across the six models and whether those differences were consistent across the six crash attributes. The repeated-measures analysis confirmed that overall agreement varied significantly among models, F(2.74, 11,482) = 1,930.8, p < 0.001, … view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

30 extracted references · 1 canonical work pages

  1. [1]

    Investigating the nature and impact of reporting bias in road crash data,

    K. A. Abay, “Investigating the nature and impact of reporting bias in road crash data,” Transportation research part A: policy and practice, vol. 71, pp. 31–45, 2015

    2015

  2. [2]

    Under-reporting of motor vehicle traffic crash victims in New Zealand,

    J. Alsop and J. Langley, “Under-reporting of motor vehicle traffic crash victims in New Zealand,” Accident Analysis & Prevention, vol. 33, no. 3, pp. 353–359, 2001

    2001

  3. [3]

    Under-reporting of road crash casualties in France,

    E. Amoros, J.-L. Martin, and B. Laumon, “Under-reporting of road crash casualties in France,” Accident Analysis & Prevention, vol. 38, no. 4, pp. 627–635, 2006

    2006

  4. [4]

    Crash data quality for road safety research: Current state and future directions,

    M. Imprialou and M. Quddus, “Crash data quality for road safety research: Current state and future directions,” Accident Analysis & Prevention, vol. 130, pp. 84–90, 2019

    2019

  5. [5]

    Understanding traffic crash under-reporting: linking police and medical records to individual and crash characteristics,

    K. H. Janstrup, S. Kaplan, T. Hels, J. Lauritsen, and C. G. Prato, “Understanding traffic crash under-reporting: linking police and medical records to individual and crash characteristics,” Traffic injury prevention, vol. 17, no. 6, pp. 580–584, 2016

    2016

  6. [6]

    Protected Bike Lanes Are Not Created Equal: A Review of Design, Safety, and Mobility Outcomes,

    S. Bharati, P. Kafle, and S. Khatiwada, “Protected Bike Lanes Are Not Created Equal: A Review of Design, Safety, and Mobility Outcomes,” Jul. 21, 2026, Zenodo. doi: 10.5281/ZENODO.21464508

    work page doi:10.5281/zenodo.21464508 2026

  7. [7]

    Semantic search approach for information extraction from traffic crash narratives,

    C. Arteaga, U. Ray, J. Park, and A. Paz, “Semantic search approach for information extraction from traffic crash narratives,” Transportation Research Interdisciplinary Perspectives, vol. 36, p. 101888, 2026

    2026

  8. [8]

    Application of text mining techniques to identify actual wrong-way driving (WWD) crashes in police reports,

    P. Hosseini, S. Khoshsirat, M. Jalayer, S. Das, and H. Zhou, “Application of text mining techniques to identify actual wrong-way driving (WWD) crashes in police reports,” International journal of transportation science and technology, vol. 12, no. 4, pp. 1038–1051, 2023

    2023

  9. [9]

    Identification and analysis of misclassified work-zone crashes using text mining techniques,

    M. A. Sayed, X. Qin, R. J. Kate, D. Anisuzzaman, and Z. Yu, “Identification and analysis of misclassified work-zone crashes using text mining techniques,” Accident Analysis & Prevention, vol. 159, p. 106211, 2021

    2021

  10. [10]

    Natural language understanding in road accident data analysis,

    J. Wu and B. Heydecker, “Natural language understanding in road accident data analysis,” Advances in Engineering Software, vol. 29, no. 7–9, pp. 599–610, 1998

    1998

  11. [11]

    Identifying secondary crashes using text mining techniques,

    X. Zhang, E. Green, M. Chen, and R. R. Souleyrette, “Identifying secondary crashes using text mining techniques,” Journal of Transportation Safety & Security, vol. 12, no. 10, pp. 1338–1358, 2020

    2020

  12. [12]

    Large language models and their applications in roadway safety and mobility enhancement: A comprehensive review,

    M. M. Karim, Y. Shi, S. Zhang, B. Wang, M. Nasri, and Y. Wang, “Large language models and their applications in roadway safety and mobility enhancement: A comprehensive review,” Artificial Intelligence for Transportation, vol. 1, p. 100004, 2025

    2025

  13. [13]

    Leveraging Large Language Models for advanced analysis of crash narratives in traffic safety research.,

    X. Wang and M. Carlino, “Leveraging Large Language Models for advanced analysis of crash narratives in traffic safety research.,” 2025

    2025

  14. [14]

    SafeTraffic Copilot: adapting large language models for trustworthy traffic safety assessments and decision interventions,

    Y. Zhao, P. Wang, Y. Zhao, H. Du, and H. F. Yang, “SafeTraffic Copilot: adapting large language models for trustworthy traffic safety assessments and decision interventions,” Nature Communications, vol. 16, no. 1, p. 8846, 2025

    2025

  15. [15]

    Injury severity on traffic crashes: A text mining with an interpretable machine-learning approach,

    C. Arteaga, A. Paz, and J. Park, “Injury severity on traffic crashes: A text mining with an interpretable machine-learning approach,” Safety Science, vol. 132, p. 104988, 2020

    2020

  16. [16]

    Topic models from crash narrative reports of motorcycle crash causation study,

    S. Das, A. Dutta, and I. Tsapakis, “Topic models from crash narrative reports of motorcycle crash causation study,” Transportation research record, vol. 2675, no. 9, pp. 449–462, 2021

    2021

  17. [17]

    Industrial benchmarking of LLMs: Assessing hallucination in traffic incident scenarios with a novel spatio-temporal dataset,

    Q. Li et al., “Industrial benchmarking of LLMs: Assessing hallucination in traffic incident scenarios with a novel spatio-temporal dataset,” 2024

    2024

  18. [18]

    Identifying the causes of road traffic collisions: Using police officers’ expertise to improve the reporting of contributory factors data,

    J. J. Rolison, “Identifying the causes of road traffic collisions: Using police officers’ expertise to improve the reporting of contributory factors data,” Accident Analysis & Prevention, vol. 135, p. 105390, 2020

    2020

  19. [19]

    Large language models for traffic and transportation research: Methodologies, state of the art, and future opportunities,

    Y. Yan et al., “Large language models for traffic and transportation research: Methodologies, state of the art, and future opportunities,” arXiv preprint arXiv:2503.21330, 2025

    Pith/arXiv arXiv 2025

  20. [20]

    Characteristics and availability of fatal road-crash databases in 20 countries worldwide,

    J. Luoma and M. Sivak, “Characteristics and availability of fatal road-crash databases in 20 countries worldwide,” Journal of safety research, vol. 38, no. 3, pp. 323–327, 2007

    2007

  21. [21]

    Under-reporting of road traffic crash data in Ghana,

    M. Salifu and W. Ackaah, “Under-reporting of road traffic crash data in Ghana,” International journal of injury control and safety promotion, vol. 19, no. 4, pp. 331–339, 2012

    2012

  22. [22]

    Crash narrative classification: Identifying agricultural crashes using machine learning with curated keywords,

    J. Kim, A. B. Trueblood, H.-C. Kum, and E. M. Shipp, “Crash narrative classification: Identifying agricultural crashes using machine learning with curated keywords,” Traffic injury prevention, vol. 22, no. 1, pp. 74–78, 2021

    2021

  23. [23]

    Identification of Factors Contributing to Traffic Crashes by Analysis of Text Narratives,

    C. D. Arteaga-Sanchez, “Identification of Factors Contributing to Traffic Crashes by Analysis of Text Narratives,” 2022

    2022

  24. [24]

    Predicting person-level injury severity using crash narratives: A balanced approach with roadway classification and natural language process techniques,

    M. Z. Majidi, S. Karimi, T. Wang, R. Kluger, and R. Souleyrette, “Predicting person-level injury severity using crash narratives: A balanced approach with roadway classification and natural language process techniques,” arXiv preprint arXiv:2509.07845, 2025. Bharati, KC Khatri, and Bharati

    Pith/arXiv arXiv 2025

  25. [25]

    Mining the highway-rail grade crossing crash data: A text mining approach,

    S. Soleimani, A. Mohammadi, J. Chen, and M. Leitner, “Mining the highway-rail grade crossing crash data: A text mining approach,” presented at the 2019 18th IEEE International Conference on Machine Learning and Applications (ICMLA), IEEE, 2019, pp. 1063–1068

    2019

  26. [26]

    Exploring traffic crash narratives in jordan using text mining analytics,

    S. Jaradat, T. I. Alhadidi, H. I. Ashqar, A. Hossain, and M. Elhenawy, “Exploring traffic crash narratives in jordan using text mining analytics,” presented at the 2024 IEEE 3rd International Conference on Computing and Machine Intelligence (ICMI), IEEE, 2024, pp. 1–6

    2024

  27. [27]

    New insights into road accident analysis through the use of text mining methods,

    S. Krause and F. Busch, “New insights into road accident analysis through the use of text mining methods,” presented at the 2019 6th International Conference on Models and Technologies for Intelligent Transportation Systems (MT-ITS), IEEE, 2019, pp. 1–6

    2019

  28. [28]

    Discovering latent themes in traffic fatal crash narratives using text mining analytics and network topology,

    K. M. Kwayu, V. Kwigizile, K. Lee, and J.-S. Oh, “Discovering latent themes in traffic fatal crash narratives using text mining analytics and network topology,” Accident Analysis & Prevention, vol. 150, p. 105899, 2021

    2021

  29. [29]

    Extracting information from narratives: An application to aviation safety reports,

    C. Posse, B. Matzke, C. Anderson, A. Brothers, M. Matzke, and T. Ferryman, “Extracting information from narratives: An application to aviation safety reports,” presented at the 2005 IEEE Aerospace Conference, IEEE, 2005, pp. 3678–3690

    2005

  30. [30]

    Towards Reliable and Interpretable Traffic Crash Pattern Prediction and Safety Interventions Using Customized Large Language Models,

    Y. Zhao, P. Wang, Y. Zhao, H. Du, and H. F. Yang, “Towards Reliable and Interpretable Traffic Crash Pattern Prediction and Safety Interventions Using Customized Large Language Models,” arXiv preprint arXiv:2505.12545, 2025

    Pith/arXiv arXiv 2025