Pith. sign in

REVIEW 4 major objections 5 minor 28 references

Enhancing Traffic Accident Classifications: Application of NLP Methods for City Safety

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read The paper claims that free-text accident descriptions dominate structured data in Munich accident classification, and that most accidents labeled 'other' are actually parking accidents.

desk verdict Useful applied NLP on Munich accident data with a plausible text-dominance claim, but the 236-sample expert test needs explicit train/test disjointness and reliability reporting. read the letter →

arxiv 2506.12092 v1 pith:7XLJNVLD submitted 2025-06-11 cs.CL cs.LG

Enes Özeren , Alexander Ulbrich , Sascha Filimon , David Rügamer , Andreas Bender This is my paper
classification cs.CLcs.LG MSC 68T5062H3068T07
keywords trafficaccidentclassificationnaturallanguageprocessingfew-shotlearningtopicmodelingXLM-Rlabelqualitymultimodalcitysafety
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper analyzes over 105,000 Munich traffic accidents to see whether the official police labels (seven types, A1–A7) can be trusted and whether they can be predicted automatically. It finds that a large share of accidents dumped into the fallback category 'other' are actually parking accidents: topic modeling groups them with stationary-vehicle accidents, and a few-shot LLM labels 69% of human A7 cases as A5 (stationary). For prediction, a fine-tuned XLM-R text model reaches 72% accuracy on an expert-labeled test set, while a tabular-only MLP reaches 53%, and a multimodal model adds little over text alone. The authors conclude that free-text descriptions are the most informative signal and that cleaning labels via human–LLM agreement produces models as good as or better than using all noisy labels.

What carries the argument

Three tools carry the argument. BERTopic, a neural topic model that embeds texts and clusters them with HDBSCAN, generates 18 topics whose cosine similarity to the seven human-determined classes shows A7 aligning nearly exactly with parking topics. Gemma-2-27B, prompted with a task description and six verified example accidents, provides a few-shot second opinion on the full dataset; its confusion matrix against human labels exposes the A7-to-A5 disagreement. XLM-RoBERTa-Large, a 560-million-parameter multilingual encoder fine-tuned on the German accident texts, is the model that demonstrates text dominance, while the multimodal variant appends a tabular MLP branch whose contribution is marginal.

What would settle it

Check whether any of the 236 expert-labeled accidents appear in the 62,000 or 45,000 training instances; if they do, remove them and re-evaluate the fine-tuned XLM-R on a strictly held-out expert set. Should accuracy fall to the tabular-only level, the text-dominance conclusion would be an artifact of label leakage rather than language understanding.

Watch

Extended reading notes

Core claim

On its own terms, the paper establishes two central findings. First, in this dataset the free-text accident description is the dominant predictive signal: a fine-tuned XLM-R model reaches 72% accuracy (0.68–0.70 weighted F1) on a 236-accident expert-labeled test set, the tabular-only MLP reaches 53% (0.49 F1), and the multimodal model barely improves on text alone (73% accuracy). Second, the fallback label A7 ('other accident'), assigned to nearly half of Munich accidents, is largely a mislabeling of parking accidents: BERTopic cluster similarities place A7 texts close to parking topics, and the few-shot Gemma-2 model reclassifies 69% of human A7 labels as A5 (stationary accident). The paper also finds that training on a smaller set of human–LLM agreed labels performs comparably to training on a larger set of presumed low-quality labels, indicating that label quality can compensate for dataset size.

Load-bearing premise

The accuracy and text-dominance claims rest on 236 expert labels treated as clean ground truth, but the paper does not report how many experts labeled each accident, whether they agreed, or whether those 236 accidents were excluded from the training data.

Editorial extensions

If this is right

  • Cities that already store free-text accident descriptions can build accurate classifiers without waiting for enriched structured data.
  • Reclassifying most A7 accidents as parking accidents turns an unknown, heterogeneous fallback category into a concrete target for countermeasures such as signage and loading-zone planning.
  • Training on a smaller set of high-confidence human–LLM agreed labels matches training on the full noisy set, so label-cleaning pipelines can improve models at no extra annotation cost in settings where few-shot labeling already runs.
  • The low overall human–LLM agreement (44%) implies systematic noise in police labels, so safety analytics built on raw labels should validate classifications before drawing policy conclusions.
  • Deploying a model as a real-time suggestion system for human labelers, with active learning, is a direct next step the paper sketches for improving label quality over time.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If the Munich findings generalize, any city with free-text police reports can run the same topic-model-plus-few-shot audit to locate systematic mislabeling before doing predictive modeling; this is a testable transfer to, say, Berlin's lower A7 rate.
  • The one-percentage-point multimodal gain suggests tabular features are mostly redundant given the text; a follow-up could test whether location data specifically adds spatial information that text alone cannot convey, which would matter for targeted safety interventions.
  • The 69% A7-to-A5 reclassification rate is only meaningful if the LLM's judgment is sound; comparing few-shot labels directly against the 236 expert labels would reveal whether the LLM or the police labels are closer to ground truth for fallback cases.
  • The comparable low- and high-quality training results imply substantial label noise; a random re-labeling study could quantify the noise rate and tell cities how much investment in label cleaning is worthwhile.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper analyzes traffic accident records from Munich (105,217 accidents, 102,569 with free-text descriptions) to identify patterns across seven accident types (A1–A7). It applies BERTopic for semantic clustering, few-shot prompting with Gemma-2-27B-Instruct to generate second-opinion labels, and supervised models (tabular MLP, fine-tuned XLM-R, and a multimodal text-plus-tabular model) for classification. The central empirical claims are (i) that free-text descriptions carry the most relevant information for accident classification, with the text-only XLM-R reaching 72% accuracy on a 236-example expert-labeled test set versus 53% for the tabular-only MLP, and (ii) that most accidents labeled as A7 (other) are actually parking-related, supported by BERTopic cluster similarities and LLM few-shot labeling. The authors also propose two training-set construction strategies (low-quality and high-quality labels) based on human–LLM agreement and report that the smaller high-quality set performs comparably to the larger low-quality set.

Significance. If the text-dominance result holds, it provides a concrete, policy-relevant demonstration that free-text police reports are more informative than structured tabular features for accident-type classification, and that fine-tuned encoder-only transformer models outperform a much larger few-shot LLM on this task. The paper also makes a practical contribution by showing that the A7 fallback category is largely composed of parking-related accidents, which can inform city planning. Strengths include public code, the use of an expert-labeled test set rather than relying solely on noisy police labels, and the combination of unsupervised topic modeling with few-shot labeling to probe label quality. However, the evaluation is built on a small test set whose construction and independence from the training set are not documented, and the label-quality analysis contains an internal tension regarding the reliability of the LLM labels used to define training-set quality.

major comments (4)
  1. [§5.1 Test Set; Table 4] The 236 expert-labeled test set is described only as 'carefully labeled' by domain specialists; the paper does not state how these accidents were selected, how many experts labeled each accident, what instructions they received, or what inter-annotator agreement was. Critically, there is no explicit statement that these 236 accidents were excluded from the 62k low-quality and 45k high-quality training sets. Since the XLM-R model is fine-tuned for 6 epochs on texts from the same corpus, any overlap between test and training texts would allow memorization of text-to-label pairs, and the reported 72% vs. 53% accuracy gap would not establish generalization from text. This must be clarified or corrected before the text-dominance claim can be accepted.
  2. [§5.1 Training Set vs. §4.2] The 'high-quality' training set is defined as those labels on which human annotators and the LLM agree, for all categories A1–A6 and agreed A7 labels. However, Section 4.2 reports that the LLM misclassifies 49% of human-labeled A2 cases as A3 and 21% of human-labeled A6 cases as A3, and the authors state that human judgment is more reliable in these cases. Using human–LLM agreement as a label-quality filter for exactly these categories is therefore internally inconsistent: it would preferentially exclude human labels that the LLM disagrees with, even though the LLM is acknowledged to be less accurate than humans on those distinctions. This biases the 'high-quality' training distribution and undermines the label-quality comparison.
  3. [§5.1 Models & Training Details] All three supervised models are evaluated by selecting the checkpoint with the 'lowest validation loss', yet no validation split is described anywhere in the manuscript. Without knowing how the validation set was constructed, whether it was disjoint from the training set, or how it relates to the 236-expert test set, the model-selection procedure is not reproducible and the reported test-set numbers cannot be independently verified.
  4. [§5.2, Table 4] The test set contains only 236 samples, and no confidence intervals or significance tests are reported for the accuracy or weighted F1 differences. In particular, the claim that the multimodal model provides only marginal improvements over text-only (0.73 vs. 0.72 on low-quality, and 0.70 vs. 0.72 on high-quality training labels) is based on differences that are well within plausible sampling error for this sample size. A bootstrap or exact binomial interval would make the extent of the tabular contribution, or lack thereof, assessable.
minor comments (5)
  1. [§1] The word 'missclassification' appears in the sentence 'we gain insights about missclassification'; it should be 'misclassification'.
  2. [§3.2] The sentence 'This finding helps reduce uncertainty about accident characteristics in the fallback category A7 ... (which is in line with our findings from Section 4.2)' refers to topic-modeling results and should cite Section 3.2, not Section 4.2.
  3. [§3.1] The embedding model is named 'jiina-embeddings-v3'; the correct spelling is 'jina-embeddings-v3'.
  4. [§5.1, Figure 4a/4b] The confusion matrices displayed for low-quality and high-quality labels in Figures 4a and 4b appear identical in the text; if these panels are intended to show which training cells are retained, the figures need to be redrawn so the 62k and 45k selection schemes are visually distinct.
  5. [§4.2] The description of the few-shot prompt states that six example descriptions with verified labels were selected per category, but it does not specify how these examples were chosen, what 'verified' means, or whether the same examples were used in the training-set construction; adding this detail would improve reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central text-vs-tabular comparison is evaluated on an external expert-labeled test set, not on the paper's own fitted labels.

full rationale

The paper's load-bearing claim, that textual data carries the most relevant information for accident classification, rests on Table 4, which reports accuracy on 236 accidents labeled by domain specialists from the City of Munich. The paper states in Section 5.1: 'We use these expert-labeled examples as our test set to report results.' This is an external, independently constructed ground truth, so the 72% vs 53% text-versus-tabular gap is not defined in terms of the models' training labels or the LLM whose outputs are used to construct training sets. Similarly, the secondary claim that most A7 (other) accidents are parking-related is supported by two independent analyses: unsupervised BERTopic clustering (Section 3.2) and Gemma few-shot labeling (Section 4.2), neither of which reduces by construction to the claim. The training-label strategies in Section 5.1 define 'high-quality' labels as human-LLM agreement, but the subsequent conclusion about label quality is an interpretation of expert-label performance, not an equation equating the definition with the result. There are no load-bearing self-citations, no fitted parameters renamed as predictions, and no uniqueness theorems imported from the authors' prior work. The reviewer's concern that the 236 test accidents might overlap with the training pool is a data-integrity and validity risk, not an instance of circular reasoning under the defined taxonomy, since the paper's own evaluation protocol is external to its modeling assumptions. The derivation chain is therefore self-contained, and no circular step can be concretely exhibited.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The study relies on several modeling choices and domain assumptions, but introduces no new theoretical entities. The main uncertainty is the assumption that the expert-labeled test set is independent and reliable.

free parameters (5)
  • UMAP hyperparameters (n_neighbors, min_dist, n_components) = not reported
    Section 3.1 lists four UMAP hyperparameters but does not state the values used for the topic-modeling results.
  • HDBSCAN min_cluster_size = not reported
    Section 3.1 mentions a minimal cluster size hyperparameter but gives no value or selection procedure.
  • Few-shot example selection = 6 examples for each of A1-A6, none for A7
    Section 4.2: six exemplary accident descriptions for each non-fallback type are hand-picked; the choice affects LLM few-shot labels.
  • Random subset size for topic model = 50,000
    Section 3.2: analysis was limited to a random subset of 50,000 accident descriptions; the number is arbitrary and not justified.
  • Label-quality agreement threshold = exact human-LLM agreement
    Section 5.1 defines high-quality labels as cases where human and LLM few-shot labels agree; no alternative thresholds are explored.
assumptions (4)
  • domain assumption Expert labels are ground truth
    Section 5.1 states 236 expert-labeled accidents are used as the test set without describing the labeling protocol or inter-rater reliability.
  • domain assumption Test accidents are independent of training data
    The paper does not state that the 236 expert-labeled accidents were excluded from the training set; without this, the evaluation could be contaminated.
  • domain assumption German text descriptions carry enough information to classify accident type
    The whole predictive approach assumes the free-text accounts are sufficient for classification; the paper does not quantify what information is missing from text.
  • domain assumption Selected models are suitable for German accident text
    Sections 3.1 and 4.1 choose jina-embeddings-v3 and Gemma-2-27B based on German benchmarks and context window requirements; their suitability for this specific domain is assumed.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Enhancing Traffic Accident Classifications: Application of NLP Methods for City Safety." pith.science (2026). https://pith.science/paper/7XLJNVLD

@misc{pith2026250612092,
  author       = {Pith},
  title        = {Pith review of: Enhancing Traffic Accident Classifications: Application of NLP Methods for City Safety},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7XLJNVLD}},
  note         = {Machine review of arXiv:2506.12092}
}
read the original abstract

A comprehensive understanding of traffic accidents is essential for improving city safety and informing policy decisions. In this study, we analyze traffic incidents in Munich to identify patterns and characteristics that distinguish different types of accidents. The dataset consists of both structured tabular features, such as location, time, and weather conditions, as well as unstructured free-text descriptions detailing the circumstances of each accident. Each incident is categorized into one of seven predefined classes. To assess the reliability of these labels, we apply NLP methods, including topic modeling and few-shot learning, which reveal inconsistencies in the labeling process. These findings highlight potential ambiguities in accident classification and motivate a refined predictive approach. Building on these insights, we develop a classification model that achieves high accuracy in assigning accidents to their respective categories. Our results demonstrate that textual descriptions contain the most informative features for classification, while the inclusion of tabular data provides only marginal improvements. These findings emphasize the critical role of free-text data in accident analysis and highlight the potential of transformer-based models in improving classification reliability.

Figures

Figures reproduced from arXiv: 2506.12092 by the authors.

Figure 1
Figure 1. First, we note that the lowest cosine similarity value is around 0.7, indicating generally high similarity, as this measure ranges up to a maximum of 1. This might be due to the fact that all texts are similar in the way that they all deal with traffic accidents. What can also be seen, for instance, is a high similarity between accident type A1 (driving accident), and the bicycle topic from the topic model (column 1… view at source ↗
Figure 1
Figure 1. Cosine similarity between BERTopic-generated and human-determined clusters. includes accident type A5 (stationary accidents), a relatively high similarity to all six extracted parking topics is visible. Finally, column seven, which represents the fallback category A7 (other accidents) shows almost exactly the same color pattern as column five. In other words, accidents labeled as “other accident” seem to be similar … view at source ↗
Figure 2
Figure 2. Few-shot prompt for classifying accidents. Some components of the prompt shortened for illustration. The last part (text to classify) is changed for each accident text and inference is performed with the Gemma-2-27B-Instruct model. This is the translated English version; the original prompt is in German since accident texts are also in German. accidents consistently contain parking-related keywords such as ’garage’,… view at source ↗
Figures from the paper (2 more)
Figure 3
Figure 3. Figure 3: Distribution of LLM few-shot labels vs. human labels. Each column sums to 100%, and bubble sizes correspond to the intersection ratio between LLM few-shot and human labels. The anti-diagonal represents the agreement ratio between the two. Since both tabular and text da…
Figure 4
Figure 4. Figure 4: Two strategies for constructing the training set. The accidents used to create the training set are highlighted in red. (a) Assuming all human-labeled instances for types A1–A6 are correct, while for type A7, only labels agreed upon by both humans and the LLM few-shot …

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

28 extracted references · 14 canonical work pages

  1. [1]

    Communications of the ACM55(4), 77–84 (2012)

    Blei, D.M.: Probabilistic topic models. Communications of the ACM55(4), 77–84 (2012)

    work page 2012

  2. [2]

    Advances in neural information processing systems 33, 1877–1901 (2020)

    Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems 33, 1877–1901 (2020)

    2020

  3. [3]

    arXiv preprint arXiv:1911.02116 (2019)

    Conneau, A., Khandelwal, K., Goyal, N., Chaudhary, V., Wenzek, G., Guzmán, F., Grave, E., Ott, M., Zettlemoyer, L., Stoyanov, V.: Un- supervised cross-lingual representation learning at scale. arXiv preprint arXiv:1911.02116 (2019)

    arXiv 2019

  4. [4]

    In: Proceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technologies, volume 1 (long and short papers)

    Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. In: Proceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technologies, volume 1 (long and short papers). pp. 4171–4186 (2019)

    2019

  5. [5]

    arXiv preprint arXiv:2203.05794 (2022)

    Grootendorst, M.: Bertopic: Neural topic modeling with a class-based tf-idf procedure. arXiv preprint arXiv:2203.05794 (2022)

    arXiv 2022

  6. [6]

    Accident Analysis & Prevention 191, 107217 (2023)

    Hossain, M.M., Zhou, H., Das, S.: Data mining approach to explore emer- gency vehicle crash patterns: A comparative study of crash severity in emer- gency and non-emergency response modes. Accident Analysis & Prevention 191, 107217 (2023)

    work page 2023

  7. [7]

    In: ICML

    Joachims, T., et al.: A probabilistic analysis of the rocchio algorithm with tfidf for text categorization. In: ICML. vol. 97, pp. 143–151. Citeseer (1997)

    work page 1997

  8. [8]

    arXiv preprint arXiv:2405.03066 (2024)

    Li, L., Zhou, J., Gao, Z., Hua, W., Fan, L., Yu, H., Hagen, L., Zhang, Y., Assimes, T.L., Hemphill, L., et al.: A scoping review of using large language models (llms) to investigate electronic health records (ehrs). arXiv preprint arXiv:2405.03066 (2024)

    arXiv 2024

Show all 28 references
  1. [9]

    McInnes, L., Healy, J., Astels, S., et al.: hdbscan: Hierarchical density based clustering. J. Open Source Softw.2(11), 205 (2017)

    2017

  2. [10]

    arXiv preprint arXiv:1802.03426 (2018)

    McInnes, L., Healy, J., Melville, J.: Umap: Uniform manifold approximation and projection for dimension reduction. arXiv preprint arXiv:1802.03426 (2018)

    2018 arXiv

  3. [11]

    arXiv preprint arXiv:2210.07316 (2022)

    Muennighoff, N., Tazi, N., Magne, L., Reimers, N.: Mteb: Massive text embedding benchmark. arXiv preprint arXiv:2210.07316 (2022). https://doi.org/10.48550/ARXIV.2210.07316, https://arxiv.org/abs/2210.07316

    2022 doi arXiv

  4. [12]

    Transportation research record2678(2), 212–222 (2024)

    Nassereddine, H., Santiago-Chaparro, K.R., Noyce, D.A.: Evaluating right- turn flashing yellow arrow for vehicle–pedestrian interactions using a non- probabilistic regression approach. Transportation research record2678(2), 212–222 (2024)

    2024

  5. [13]

    In: 2020 International joint conference on neural networks (IJCNN)

    Noguti, M.Y., Vellasques, E., Oliveira, L.S.: Legal document classification: An application to law area prediction of petitions to public prosecution Enhancing Traffic Accident Classifications 17 service. In: 2020 International joint conference on neural networks (IJCNN). pp. ...

    2020

  6. [14]

    Sensors21(20), 6929 (2021)

    Park, S., Park, S., Jeong, H., Yun, I., So, J.: Scenario-mining for level 4 automated vehicle safety assessment from real accident situations in urban areas using a natural language process. Sensors21(20), 6929 (2021)

    2021

  7. [15]

    Information14(4), 242 (2023)

    Patwardhan, N., Marrone, S., Sansone, C.: Transformers in the real world: A survey on nlp applications. Information14(4), 242 (2023)

    2023

  8. [16]

    Radford, A., Narasimhan, K., Salimans, T., Sutskever, I., et al.: Improving language understanding by generative pre-training (2018)

    2018

  9. [17]

    Journal of machine learning research21(140), 1– 67 (2020)

    Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., Zhou, Y., Li, W., Liu, P.J.: Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research21(140), 1– 67 (2020)

    2020

  10. [18]

    In: Inui, K., Jiang, J., Ng, V., Wan, X

    Reimers, N., Gurevych, I.: Sentence-BERT: Sentence embeddings using Siamese BERT-networks. In: Inui, K., Jiang, J., Ng, V., Wan, X. (eds.) Proceedings of the 2019 Conference on Empiri- cal Methods in Natural Language Processing and the 9th Interna- tional Joint Conference on N...

    2019 doi

  11. [19]

    IEEE Open Journal of Intelligent Transportation Systems4, 955–965 (2023)

    Seo, Y., Park, J., Oh, G., Kim, H., Hu, J., So, J.: Text classification model- ing approach on imbalanced-unstructured traffic accident descriptions data. IEEE Open Journal of Intelligent Transportation Systems4, 955–965 (2023)

    2023

  12. [20]

    Siino, M., Falco, M., Croce, D., Rosso, P.: Exploring llms applications in law:Aliteraturereviewoncurrentlegalnlpapproaches.IEEEAccess(2025)

    2025

  13. [21]

    Sturua, S., Mohr, I., Akram, M.K., Günther, M., Wang, B., Krimmel, M., Wang, F., Mastrapas, G., Koukounas, A., Koukounas, A., Wang, N., Xiao, H.: jina-embeddings-v3: Multilingual embeddings with task lora (2024), https://arxiv.org/abs/2409.10173

    2024 arXiv

  14. [22]

    arXiv preprint arXiv:2408.00118 (2024)

    Team, G., Riviere, M., Pathak, S., Sessa, P.G., Hardin, C., Bhupatiraju, S., Hussenot, L., Mesnard, T., Shahriari, B., Ramé, A., et al.: Gemma 2: Improving open language models at a practical size. arXiv preprint arXiv:2408.00118 (2024)

    2024 arXiv

  15. [23]

    Thellmann, K., Stadler, B., Fromm, M., Buschhoff, J.S., Jude, A., Barth, F., Leveling, J., Flores-Herr, N., Köhler, J., Jäkel, R., et al.: Towards multilin- gualllmevaluationforeuropeanlanguages.arXivpreprintarXiv:2410.08928 (2024)

    2024 arXiv

  16. [24]

    Advances in neural information processing systems30 (2017)

    Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, Ł., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems30 (2017)

    2017

  17. [25]

    Advances in neural information processing systems35, 24824–24837 (2022) 18 E

    Wei, J., Wang, X., Schuurmans, D., Bosma, M., Xia, F., Chi, E., Le, Q.V., Zhou, D., et al.: Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems35, 24824–24837 (2022) 18 E. Özeren and A. Ulbrich et al

    2022

  18. [26]

    arXiv preprint arXiv:1910.03771 (2019)

    Wolf, T., Debut, L., Sanh, V., Chaumond, J., Delangue, C., Moi, A., Cistac, P., Rault, T., Louf, R., Funtowicz, M., et al.: Huggingface’s transformers: State-of-the-art natural language processing. arXiv preprint arXiv:1910.03771 (2019)

    2019 arXiv

  19. [27]

    Advances in Engineering Software29(7-9), 599–610 (1998)

    Wu, J., Heydecker, B.: Natural language understanding in road accident data analysis. Advances in Engineering Software29(7-9), 599–610 (1998)

    1998

  20. [28]

    Applied Sciences14(17), 7819 (2024)

    Xing, X., Chen, P.: Entity extraction of key elements in 110 police reports based on large language models. Applied Sciences14(17), 7819 (2024)

    2024

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.