Pith. sign in

REVIEW 2 major objections 4 minor 23 references

Utilizing Multimodal Data for Edge Case Robust Call-sign Recognition and Understanding

T0 review · 2 major / 4 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read A multimodal model recovers call-signs when transcripts are noisy, clipped, or missing entirely, adding up to 15% edge-case accuracy over text-only models.

desk verdict Solid ATC speech paper with two new architectures; the CallSBERT efficiency result is credible, but the CCR edge-case gains hinge on an unstated train/test split for the map-building data that a referee must pin down. read the letter →

arxiv 2412.20467 v1 pith:BKTOC3HD submitted 2024-12-29 cs.CL

Alexander Blatt , Dietrich Klakow This is my paper · ORCID
classification cs.CL
keywords call-signrecognitionairtrafficcontrolmultimodalfusionedgecasesnoiserobustnesscommanddistributioncontrastivelearningsurveillancedata
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

This paper targets air-traffic control call-sign recognition and understanding (CRU), the task of identifying which aircraft a controller's utterance addresses. It argues that a multimodal model combining the transcript with surveillance data—plane coordinates and predicted command types—can keep call-sign accuracy high in edge cases where pure text models collapse: high word error rates, clipped utterances, and even completely missing transcripts. The authors introduce two models: CallSBERT, a smaller and faster contrastively trained CRU model, and the call-sign-command recovery (CCR) architecture that wraps CallSBERT with a command-distribution branch. They report up to 15% absolute accuracy gains in edge cases, and show that on transcript-free inputs CCR retains roughly 37–40% accuracy where text-only models fall to near zero. The claim matters because operational ATC assistant systems face exactly these degraded conditions, and a model that can fall back on surveillance context is safer and more deployable.

What carries the argument

The load-bearing component is the command distribution module (CDM), which maps plane coordinates to command probabilities for six command types (horizontal, vertical, ils, taxi, clearing, greeting). The CDM is built from a small set of coordinate-command pairs of the target airspace, filtered with a Gaussian function to produce dense probability maps over the 200 km × 200 km × 20 km airspace volume. A transformer-based command classifier selects the relevant command distribution(s) for a given transcript; when no transcript is available, all distributions are mean-pooled. The final call-sign identifier is a five-layer fully connected network that combines CallSBERT's per-call-sign similarity scores (Sim) with the CDM's per-call-sign distribution scores (Dis) to produce a weighted score and extract the most probable call-sign. The Gaussian filtering allows probability maps to be generated from as few as 100 coordinate-command pairs, which is what lets the system generalize to a new airspace with little annotated data.

What would settle it

Rebuild the CDM using coordinate-command pairs drawn only from a time window strictly before the test transcripts (or from randomly shuffled coordinates), retrain the final identifier, and compare transcript-free and high-WER call-sign accuracy. If the CCR advantage disappears or drops below 10%, the reported edge-case gains are partly artifacts of label leakage rather than genuine multimodal generalization.

Watch

Extended reading notes

Core claim

The central claim is that optimizing CRU for edge cases—not just for clean transcripts—produces a model that is both more robust and no worse on standard conditions, and that incorporating multimodal surveillance context (coordinates and command distributions) is what makes this possible. Specifically, the paper shows that a CRU model trained on high-WER and clipped transcripts suffers far less degradation at test time, and that the CCR architecture, which fuses CallSBERT's cosine-similarity scores with command-distribution scores derived from plane positions, maintains call-sign accuracy even when the transcript is entirely absent. The paper also demonstrates that CallSBERT, with 24.6M parameters versus the 66.3M of the prior EncDec model, trains about four times faster, is robust to the number of surveillance call-signs seen during fine-tuning, and achieves comparable or better accuracy when integrated into CCR. The authors conclude that edge-case optimization plus position-based command priors yields a smaller, faster, and more robust CRU model across a wide operational range.

Load-bearing premise

The CDM probability maps are built from an unspecified small set of coordinate-command pairs of the target airspace, and the paper never states that these pairs are disjoint from the test transcripts or that the filter parameters (e.g., Gaussian bandwidth) were chosen on a separate validation set; if the map-building data overlaps the test period, the distribution scores could be leaking the true command labels into the call-sign prediction.

Editorial extensions

If this is right

  • If CCR's edge-case gains hold, operational ATC assistant systems can maintain call-sign accuracy on the roughly 10% of recordings with SNR below 0 dB that are currently discarded for ASR.
  • The finding that training on high-WER and clipped transcripts does not hurt clean-data accuracy supports a practical recipe: deliberately include degraded samples in fine-tuning to improve robustness across the operational range.
  • CallSBERT's smaller size and faster fine-tuning make it a more flexible CRU component for real-time or resource-constrained deployments, and its similarity-score output can serve as a feature for downstream modules.
  • The CCR architecture demonstrates a general template for multimodal NLU: when transcript evidence is unreliable, position- and command-based priors can stand in, keeping the system operational even when speech recognition fails completely.
  • The paper's ablation shows that combining noise-robust text features with surveillance distributions is key; neither the text branch nor the distribution branch alone matches the full CCR accuracy at high WER, implying that fusing complementary evidence is the core mechanism.

Reading between the lines

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

  • The CDM's reliance on coordinate-command pairs raises a testable leakage concern: if the map-building data overlaps the test transcripts' time period, the Dis scores could directly encode test-time command labels. Rebuilding the CDM from a disjoint time period or from synthetic coordinates would clarify whether the reported 37–40% transcript-free accuracy is genuine.
  • The same multimodal recovery idea should transfer to other domains where target identities have known positions and command-like actions are spatially structured, such as nautical vessel communications or military units; the paper hints at this but does not demonstrate it.
  • A natural extension is to fuse the CDM scores with an ASR confidence signal, so that the system can dynamically weight the text branch when ASR is confident and fall back to the distribution branch when it is not, potentially yielding smoother accuracy curves than the current fixed fusion.
  • The claim that edge-case optimization improves clean-data accuracy could be tested per-command type; if some command types (e.g., taxi) have weaker spatial priors, the robustness gains should be uneven, offering a diagnostic for where multimodal fusion actually helps.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 4 minor

Summary. The paper proposes CallSBERT, a contrastive SBERT-based call-sign recognition and understanding (CRU) model, and CCR, a multimodal architecture that augments CallSBERT with a command classifier and a command distribution module (CDM) built from ADS-B coordinates. Experiments on MALORCA and AIRBUS data claim that CCR increases edge-case call-sign accuracy by up to 15% over the text-only model, stabilizes performance under high word error rate and clipping, and retains 37-40% accuracy even when the transcript is completely missing. The authors also report that CallSBERT has fewer parameters, is faster to fine-tune, and is more robust during fine-tuning than the EncDec state-of-the-art model.

Significance. If the reported numbers survive a proper data split and parameter-selection protocol, this is a practically valuable result: a smaller CRU model that is faster to train and substantially more robust on noisy, clipped, or transcript-free inputs. The non-binary 3D command distribution maps are a sensible extension of prior 2D binary maps, and the paper includes an ablation study (Figure 6d) that isolates the contribution of the command classifier. The authors also report means and standard deviations over three runs, which is good experimental practice. However, the central evidence depends on whether the CDM coordinate-command pairs overlap the test set and whether the CDM filter parameters were selected on a validation set; the paper currently does not establish this, so the headline gains in Table 1 and Figure 6c are not yet fully supported.

major comments (2)
  1. [Section 4.3-4.4, Figure 3, Table 1] The paper never states whether the coordinate-command pairs used to build the CDM probability maps are disjoint from the test transcripts. Section 4.3 says only that the maps are generated from 'a small set of coordinate-command pairs of the target airspace,' and Section 4.4 selects the filter (Gaussian, sigma, window size) based on 'maximum accuracy' as shown in Figure 3. With a test split of only 0.1K samples, if the map-building pairs include transcripts from the test period, the Dis scores for each test call-sign directly encode the ground-truth command and plane position, which would inflate the no-transcript results in Table 1 (37-40%) and the up-to-15% edge-case gains in Section 5.2.1. The authors must specify the split used for CDM construction and filter selection, and if no split was used, they must re-run the experiments with a validation-based selection protocol.
  2. [Section 5.2.1, Figure 6d] The ablation study comparing CDM-command to CDM-naive also depends on the same provenance issue. Figure 6d reports that using the command classifier to select the relevant distribution map adds roughly 10% accuracy. If the command labels used to build the CDM maps overlap the transcripts on which the command classifier is evaluated, this 10% gap is inflated by label leakage rather than by the classifier's ability to select the correct map. The paper needs to clarify whether the CDM construction data and the command-classifier evaluation data are disjoint, and to re-measure the ablation if they are not.
minor comments (4)
  1. [Abstract and Introduction] There are several typos: 'This hold especially true' should be 'This holds especially true,' and 'ACTOs' should be 'ATCOs' in Section 1.
  2. [Section 4.2] 'one mayor drawback' should be 'one major drawback,' and 'based one SBERT block' should be 'based on an SBERT block.'
  3. [Section 5] The text says 'trancripts' in Section 5.2.2; this should be 'transcripts.'
  4. [Figure 4 caption] The caption lists panels (a), (b), and (c) but the text refers to '(b),(d)'; the panel letters and the text references should be aligned.

Circularity Check

1 steps flagged · score 4.0 of 10

CDM map construction and filter selection are not shown to be disjoint from test labels, so the CCR no-transcript and edge-case gains are vulnerable to label leakage; the CallSBERT-vs-EncDec comparison itself is independent.

  1. fitted input called prediction [Section 4.3 (CDM), Section 4.4 (CDM optimization), Section 5.2.3 / Table 1]
    "For generating the mappings, a small set of coordinate-command pairs of the target airspace are filtered by one of the following filter functions: Gaussian, binary, maximum or uniform. ... The Dis scores therefore indicate which plane in the airspace is most likely mentioned in the transcript based on its position and the command uttered in the transcript."

    The CDM maps are fit to labeled 'coordinate-command pairs' and the paper never states those pairs exclude the MALORCA test split (Section 3: 0.1K test transcripts). The CCR identifier ranks call-signs using the Dis scores, so if a test transcript's command label and plane coordinates are among the map-building pairs, that call-sign's Dis score is high by construction. The headline no-transcript results (Table 1: 0.16/0.33/0.37) and the up-to-15% edge-case gain (Abstract, Figure 6c) then reduce to looking up the training label rather than predicting.

full rationale

The paper's CallSBERT architecture and the high-WER/clipping comparisons are not circular: CallSBERT is a standard contrastive fine-tune of SBERT, and the degraded-transcript experiments use altered inputs with no encoded answer. The EncDec baseline is a prior work by the same authors, but it is used as a comparison point, not as a load-bearing uniqueness or derivation step. The central circularity risk is confined to the CDM: its probability maps are generated from a small set of coordinate-command pairs, and the paper does not state that these pairs are disjoint from the test set or that the Gaussian filter was chosen on a validation set rather than the evaluation set. As published, the missing-transcript accuracy and the up-to-15% edge-case gains are not verifiably independent predictions; if the map-building pairs overlap the test labels, those results are forced by construction. This warrants a moderate circularity score, while leaving open that a proper data split would make the multimodal contribution valid.

Assumptions & free parameters 1 free parameters · 4 assumptions · 0 invented entities

The main uncharged assumptions are domain-level: the addressee is inside the surveillance bounding box, the coordinate-command maps generalize from a small sample to the whole airspace, simulated degradations represent real operational noise, and SBERT embeddings carry enough signal for call-sign ranking. The only explicit fitted quantity is the CDM filter parameter, which appears to be tuned on the evaluation data.

free parameters (1)
  • CDM filter parameters (Gaussian sigma, window size, filter type) = not reported (selected for maximum accuracy)
    Section 4.4/Figure 3: filter type and parameters were optimized for call-sign accuracy on the evaluation data, making them fit to the test set rather than fixed a priori.
assumptions (4)
  • domain assumption The call-sign addressee is among the ~30 aircraft in the 200km x 200km x 20km bounding box around Prague airport at the transcript timestamp.
    Section 3 states roughly 30 planes are in the bounding box, so random baseline is 1/30; the whole CRU task depends on this containment assumption.
  • domain assumption The command distribution maps learned from a small set of coordinate-command pairs generalize to the test transcripts of the same airspace.
    Section 4.3 uses a small set of coordinate-command pairs filtered by Gaussian to generate probability distributions for the whole airspace; this generalization is assumed rather than demonstrated.
  • domain assumption Simulated ASR noise and clipping used to create edge-case versions approximate real operational degradation.
    Section 3 says versions of different WERs are produced by adding ASR noise as described in [8] and clipped by removing words; no real noisy ATC test set is used.
  • domain assumption SBERT sentence embeddings provide enough semantic signal to rank matching vs non-matching call-signs from transcripts.
    Section 4.2 relies on cosine similarity between the SBERT-encoded transcript and call-sign candidates without a learned scoring layer.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Utilizing Multimodal Data for Edge Case Robust Call-sign Recognition and Understanding." pith.science (2026). https://pith.science/paper/BKTOC3HD

@misc{pith2026241220467,
  author       = {Pith},
  title        = {Pith review of: Utilizing Multimodal Data for Edge Case Robust Call-sign Recognition and Understanding},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BKTOC3HD}},
  note         = {Machine review of arXiv:2412.20467}
}
read the original abstract

Operational machine-learning based assistant systems must be robust in a wide range of scenarios. This hold especially true for the air-traffic control (ATC) domain. The robustness of an architecture is particularly evident in edge cases, such as high word error rate (WER) transcripts resulting from noisy ATC recordings or partial transcripts due to clipped recordings. To increase the edge-case robustness of call-sign recognition and understanding (CRU), a core tasks in ATC speech processing, we propose the multimodal call-sign-command recovery model (CCR). The CCR architecture leads to an increase in the edge case performance of up to 15%. We demonstrate this on our second proposed architecture, CallSBERT. A CRU model that has less parameters, can be fine-tuned noticeably faster and is more robust during fine-tuning than the state of the art for CRU. Furthermore, we demonstrate that optimizing for edge cases leads to a significantly higher accuracy across a wide operational range.

Figures

Figures reproduced from arXiv: 2412.20467 by the authors.

Figure 1
Figure 1. Architecture comparison of the parallel EncDec [8] (left) and the sequential CallSBERT model (right). ils, taxi, clearing and greeting. A transcript is tagged with hor￾izontal if it contains for example the key words turn right or change heading. For each of the MALORCA transcripts, the ADS-B infor￾mation of each airplane in 100 km N-S and E-W distance of the Prague airport and 0-20 km altitude is fetched from the O… view at source ↗
Figure 5
Figure 5. Call-sign accuracy depending on the surveillance size per test transcript. During fine-tuning, each transcript has ei￾ther 4 or 24 corresponding surveillance call-signs. 5. Results 5.1. CallSBERT: Surveillance adaptation Depending on the the flight sector, the amount of surveillance call-signs available might vary. The EncDec architecture is proven to be robust against fluctuations in the surveillance call￾sign coun… view at source ↗
Figure 3
Figure 3. Maximum accuracy of call-sign prediction based on command distributions with optimal filter parameters. tween the fully connected layers and a sigmoid activation at the last layer. 4.4. CDM optimization To reduce the need for a large transcribed corpus to create the command probability distributions of the CDM for a new airspace, we evaluate different filter functions for the distribu￾tion generation in a low-resour… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: 2D coordinates of airplanes while receiving a vertical command (a) and 2D distribution maps (top view) of the vertical command in the 200 km · 200 km Prague airspace (b),(d). Dark colored areas have a high probability for vertical commands. (a) CallSBERT (b) EncDec [P…
Figure 6
Figure 6. Figure 6: Call-sign accuracy depending on the WER of the MALORCA test data. Up to a test WER of 40%, the call-sign accuracy of CallSBERTs is more than 20% higher than the CA of CDMcommand. At higher WERs the accuracy of CallSBERT drops significantly. The full CCR architecture ho…
Figure 7
Figure 7. Figure 7: Call-sign accu￾racy depending on the num￾ber of words clipped off at the beginning of the tran￾scripts [PITH_FULL_IMAGE:figures/full_fig_p004_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

23 extracted references · 20 canonical work pages

  1. [1]

    Introduction Pilots rely on the guidance of air-traffic controllers (ATCO) for a safe take-off and landing. Research projects targeting ACTO pilot communication automation, such as AcListant, Malorca [1], or ATCO2 [2], are enabling the development of fully auto- mated air traffic-control (ATC) speech processing pipelines and assistant systems. These syste...

    work page

  2. [2]

    Related work Related works focus on call-sign tagging [9], call-sign tran- scription [10] or call-sign recognition in International Civil Avi- ation Organization (ICAO) format from ATC conversation tran- scripts [8, 11]. Multimodal approaches for automatic speech recognition (ASR) in ATC use surveillance call-signs from Au- tomatic Dependent Surveillance–...

    work page

  3. [3]

    Utilizing Multimodal Data for Edge Case Robust Call-sign Recognition and Understanding

    Data preparation The CRU models are trained on ATC transcripts of the MAL- ORCA data set (Prague airport) and on transcripts of the AIR- BUS data set. Both data sets contain ATC transcripts labeled with the correct call-signs, e.g. ryanair one two four (expanded format), respectively RYR124 (ICAO format) as shown in Figure 2. The AIRBUS dataset, with arti...

    work page Pith review arXiv 2024

  4. [4]

    EncDec As SOTA, we take the EncDec model from Blatt et al

    Models 4.1. EncDec As SOTA, we take the EncDec model from Blatt et al. [8] which uses a bert-base3 encoder-decoder architecture and has 66.3M parameters. One mayor drawback of the EncDec architecture is the way, the model is trained. The model input consists of a transcript concatenated with all surveillance call-signs to pre- dict the target call-sign di...

    work page

  5. [5]

    CallSBERT: Surveillance adaptation Depending on the the flight sector, the amount of surveillance call-signs available might vary

    Results 5.1. CallSBERT: Surveillance adaptation Depending on the the flight sector, the amount of surveillance call-signs available might vary. The EncDec architecture is proven to be robust against fluctuations in the surveillance call- sign count during testing [8]. The question remains how the EncDec architecture and CallSBERT react, when they are fine...

    work page

  6. [6]

    Fine-tuning on noisy transcripts reduces the noise intro- duced accuracy drop significantly without degrading accuracy levels on clean data

    Conclusion In this work we have shown at the example of call-sign recog- nition and understanding models, that edge case optimization leads to a more stable performance over a broad operational range. Fine-tuning on noisy transcripts reduces the noise intro- duced accuracy drop significantly without degrading accuracy levels on clean data. This holds true...

    work page

  7. [7]

    Survey of Hallucination in Natural Language Generation,

    Z. Ji, N. Lee, R. Frieske, T. Yu, D. Su, Y . Xu, E. Ishii, Y . J. Bang, A. Madotto, and P. Fung, “Survey of Hallucination in Natural Language Generation,” ACM Computing Surveys, vol. 55, no. 12, 2023

    work page 2023

  8. [8]

    Iterative learning of speech recognition models for air traffic control,

    A. Srinivasamurthy, P. Motlicek, M. Singh, Y . Oualil, M. Kleinert, H. Ehr, and H. Helmke, “Iterative learning of speech recognition models for air traffic control,” Proceedings of the Annual Con- ference of the International Speech Communication Association, INTERSPEECH, vol. 2018-Septe, no. September, pp. 3519–3523, 2018

    work page 2018

Show all 23 references
  1. [9]

    Automatic call sign detection: Matching air surveillance data with air traffic spoken communications,

    J. Zuluaga-Gomez, K. Vesel ´y, A. Blatt, P. Motlicek, D. Klakow, A. Tart, I. Sz ¨oke, A. Prasad, S. Sarfjoo, P. Kol ˇc´arek, M. Kocour, H. ˇCernock´y, C. Cevenini, K. Choukri, M. Rigault, and F. Landis, “Automatic call sign detection: Matching air surveillance data with air tr...

    2020

  2. [10]

    EASA Concept Paper: First usable guidance for Level 1 machine learning applications,

    European Union Aviation Safety Agency, “EASA Concept Paper: First usable guidance for Level 1 machine learning applications,” no. 1, pp. 1–174, 2021. [Online]. Available: https://www.easa.e uropa.eu/en/easa-concept-paper-first-usable-guidance-level-1-m achine-learning-applicat...

    2021

  3. [11]

    The ATCOSIM corpus of non-prompted clean air traffic control speech,

    K. Hofbauer, S. Petrik, and H. Hering, “The ATCOSIM corpus of non-prompted clean air traffic control speech,” in Proceedings of the 6th International Conference on Language Resources and Evaluation, LREC 2008 , 2008, pp. 2147–2152. [Online]. Available: http://www.lrec-conf.org...

    2008

  4. [12]

    The Airbus Air Traffic Control Speech Recognition 2018 Challenge: Towards ATC Automatic Transcription and Call Sign Detection,

    T. Pellegrini, J. Farinas, E. Delpech, and F. Lancelot, “The Airbus Air Traffic Control Speech Recognition 2018 Challenge: Towards ATC Automatic Transcription and Call Sign Detection,” in Proc. Interspeech 2019, 2019, pp. 2993–2997

    2018

  5. [13]

    Automatic Processing Pipeline for Collecting and Annotating Air-Traffic V oice Communication Data,

    M. Kocour, K. Vesel ´y, I. Sz ¨oke, S. Kesiraju, J. Zuluaga-Gomez, A. Blatt, A. Prasad, I. Nigmatulina, P. Motl ´ıˇcek, D. Klakow, A. Tart, H. Atassi, P. Kol ˇc´arek, J. ˇCernock´y, C. Cevenini, K. Choukri, M. Rigault, F. Landis, S. Sarfjoo, and C. Salamin, “Automatic Processi...

    2022

  6. [14]

    and we adapt BERT [15] as command classifier in our edge case robust CCR architecture

  7. [15]

    Call- Sign Recognition and Understanding for Noisy Air-Traffic Tran- scripts Using Surveillance Information,

    A. Blatt, M. Kocour, K. Vesel ´y, I. Sz ¨oke, and D. Klakow, “Call- Sign Recognition and Understanding for Noisy Air-Traffic Tran- scripts Using Surveillance Information,” ICASSP , IEEE Interna- tional Conference on Acoustics, Speech and Signal Processing - Proceedings, vol. 2...

    2022

  8. [16]

    CRIM’s Speech Transcription and Call Sign Detection System for the ATC Airbus Challenge task,

    V . Gupta, L. Rebout, G. Boulianne, P. A. M ´enard, and J. Alam, “CRIM’s Speech Transcription and Call Sign Detection System for the ATC Airbus Challenge task,” inProceedings of the Annual Conference of the International Speech Communication Associ- ation, INTERSPEECH , vol. 2...

    2019

  9. [17]

    Improving callsign recognition with air-surveillance data in air-traffic communication,

    I. Nigmatulina, R. Braun, J. Zuluaga-Gomez, and P. Motlicek, “Improving callsign recognition with air-surveillance data in air-traffic communication,” aug 2021. [Online]. Available: http://arxiv.org/abs/2108.12156

    2021 arXiv

  10. [18]

    Robust Command Recognition for Lithuanian Air Traffic Control Tower Utter- ances,

    O. Ohneiser, S. Sarfjoo, H. Helmke, S. Shetty, P. Motlicek, M. Kleinert, H. Ehr, and ˇS. Murauskas, “Robust Command Recognition for Lithuanian Air Traffic Control Tower Utter- ances,” Tech. Rep., 2021

    2021

  11. [19]

    A context- aware language model to improve the speech recognition in air traffic control,

    D. Guo, Z. Zhang, P. Fan, J. Zhang, and B. Yang, “A context- aware language model to improve the speech recognition in air traffic control,” Aerospace, vol. 8, no. 11, nov 2021

    2021

  12. [20]

    Machine learning of controller command prediction models from recorded radar data and controller speech utterances,

    M. Kleinert, H. Helmke, G. Siol, H. Ehr, M. Finke, Y . Oualil, and A. Srinivasamurthy, “Machine learning of controller command prediction models from recorded radar data and controller speech utterances,” Tech. Rep., 2017. [Online]. Available: https://www.malorca-project.de/wp...

    2017

  13. [21]

    Sentence-bert: Sentence embeddings using siamese bert-networks,

    N. Reimers and I. Gurevych, “Sentence-bert: Sentence embeddings using siamese bert-networks,” in Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing. Association for Computational Linguistics, 11

    2019

  14. [23]

    BERT: pre-training of deep bidirectional transformers for language understanding,

    J. Devlin, M. Chang, K. Lee, and K. Toutanova, “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,...

    2019

  15. [2019]

    Available: https://arxiv.org/abs/1908.10084

    [Online]. Available: https://arxiv.org/abs/1908.10084

    1908 arXiv

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.