An analog-AI chip for energy-efficient speech recognition and transcription
Open accessThe DOI matches its Crossref record; 45 of the 46 references it cites were found in a registry or an index; 12 of the 36 sources its claims cite were checked for support.
Abstract
Abstract Models of artificial intelligence (AI) that have billions of parameters can achieve high accuracy across a range of tasks 1,2 , but they exacerbate the poor energy efficiency of conventional general-purpose processors, such as graphics processing units or central processing units. Analog in-memory computing (analog-AI) 3–7 can provide better energy efficiency by performing matrix–vector multiplications in parallel on ‘memory tiles’. However, analog-AI has yet to demonstrate software-equivalent (SW eq ) accuracy on models that require many such tiles and efficient communication of neural-network activations between the tiles. Here we present an analog-AI chip that combines 35 million phase-change memory devices across 34 tiles, massively parallel inter-tile communication and analog, low-power peripheral circuitry that can achieve up to 12.4 tera-operations per second per watt (TOPS/W) chip-sustained performance. We demonstrate fully end-to-end SW eq accuracy for a small keyword-spotting network and near-SW eq accuracy on the much larger MLPerf 8 recurrent neural-network transducer (RNNT), with more than 45 million weights mapped onto more than 140 million phase-change memory devices across five chips.
Overview
CoChat Summary
## TL;DR
The paper demonstrates a 14-nm analog-AI inference chip for speech recognition and transcription, using 34 tiles of phase-change memory and low-power routing to reach up to 12.4 TOPS/W chip-sustained performance. It reports end-to-end software-equivalent accuracy for keyword spotting and near-software-equivalent accuracy for the much larger MLPerf RNNT model with more than 45 million weights mapped across five chips.
## Key findings
- The chip combines 35 million phase-change memory devices across 34 tiles and achieves up to 12.4 TOPS/W chip-sustained performance.
- For keyword spotting, the system achieved 86.14% accuracy, which the paper says is within the MLPerf software-equivalent limit of 85.88% (99% of the original software model’s accuracy); the original software model is reported at 86.75%.
- For RNNT on Librispeech, the full five-chip system reached 9.475% WER in one configuration, versus a 7.452% software baseline, and with weight expansion improved to 9.258% WER.
- The RNNT mapping used 45,261,568 of 45,321,309 total network weight and bias parameters in analog memory (99.9%), spread across more than 140 million PCM devices on 5 chips.
- The paper reports a projected system performance of 6.94 TOPS/W and 546.6 samples/s/W at 3.57 W, and says this is a 14-fold improvement over the best MLPerf result for the task.
## Methods
- Experimental work used a 14-nm inference chip built from 300-mm wafers, with PCM devices added in back-end-of-line processing; experiments were run with three Xilinx Virtex-7 VC707 FPGAs and x86/MATLAB control.
- The authors tested two workloads: a keyword-spotting (KWS) model on the Google Speech Commands dataset and the MLPerf RNNT model on Librispeech.
- Main analyses included hardware-aware training, pruning, calibration, analog weight programming, and weight expansion, plus end-to-end inference measurements of accuracy/WER, power, and TOPS/W; RNNT inference was cascaded across five chips.
## Limitations
- The paper states that the 5-chip RNNT implementation is not integrated with digital processing; some digital operations are estimated or performed off-chip, so the reported system performance is partly projected rather than fully monolithic.
- It also states that the chip does not include on-chip digital computing cores or SRAM for auxiliary operations and data staging, which the text makes evident as a limitation for a marketable product.
- The paper notes that the joint-FC layer and some vector-vector, bias, and nonlinear operations were computed in software/off-chip, so not all of RNNT was done in analog hardware.
Model-written summary. Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Figures
















Verification
Claims the paper makes
4 Valid · 1 Not valid · 8 Not checked · 103 Own results and interpretations
The sources with a verdict support this claim; the rest weren’t checked yet, have no full text we could read, or their verdicts aren’t rated yet.
In the paper
“Analog in-memory computing (analog-AI)3–7 can provide better energy efficiency by performing matrix–vector multiplications in parallel on ‘memory tiles’.”
§ An analog-AI chip for energy-efficient speech recognition and transcription > Abstract
What the sources say
- Not checked — no full text of the sourcecites [3] Ambrogio, S. et al. Equivalent-accuracy accelerated neural-network training using analogue memory. Nature 558, 60–67 (2018). source
- Supportedcites [4] Narayanan, P. et al. Fully on-chip MAC at 14 nm enabled by accurate row-wise programming of PCM-based weights and parallel vector-transport in duration-format. IEEE Trans. Electron. Devices 68, 6629–6636 (2021). source
The Introduction (p. 1) describes analog memory arrays performing multiplication and accumulation where the weights are stored. It says these arrays enable massively parallel computation and avoid energy-intensive data transfers in conventional digital systems, supporting the claim's qualified energy-efficiency rationale. The paper also describes a multi-tile chip (Conclusion, p. 7). (LLM, conf=0.88)
- Supportedcites [5] Khaddam-Aljameh, R. et al. HERMES-core—a 1.59-TOPS/mm2 PCM on 14-nm CMOS in-memory compute core using 300-ps/LSB linearized CCO-based ADCs. IEEE J. Solid-State Circuits 57, 1027–1038 (2022). source
The Introduction (p. 1) says in-memory computing avoids energy-costly data movement and describes matrix–vector multiplication using an array of processing elements. Section II (p. 2) states that the PCM core performs fully parallel analog matrix–vector multiplication in its memory-cell array, supporting the claim’s description of parallel computation on memory tiles. (LLM, conf=0.91)
- Not checked — no full text of the sourcecites [6] Yao, P. et al. Fully hardware-implemented memristor convolutional neural network. Nature 577, 641–646 (2020). source
- Supportedcites [7] Wan, W. et al. A compute-in-memory chip based on resistive random-access memory. Nature 608, 504–512 (2022). source
The Main section says that highly parallel analogue computation in RRAM compute-in-memory architecture brings superior efficiency and describes in-memory matrix–vector multiplication. The Methods section describes parallel MVM across multiple cores. The paper cautions that reported peak energy efficiency excludes intermediate data movement, but that does not negate the citing sentence’s qualified claim that the approach can provide better efficiency. (LLM, conf=0.91)
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“In particular, transformer1 and recurrent neural-network transducer (RNNT)12,13,16 models containing up to one billion parameters2 have produced a marked decrease in word error rate (WER) (and therefore much better accuracy) for the automated transcription of spoken English-language sentences, as shown in Fig. 1a for two widely used datasets, Librispeech17 and SwitchBoard18.”
§ An analog-AI chip for energy-efficient speech recognition and transcription > Main
What the sources say
- Partly supportedcites [1] Vaswani, A. et al. Attention is all you need. In NIPS17: Proc. 31st Conference on Neural Information Processing Systems (eds. von Luxburg, U. et al.) 6000–6010 (Curran Associates, 2017). sourceExperimental — accuracy not yet measured
Our copy of Vaswani et al. (2017) is indeed the source of the Transformer architecture named in the citing sentence (Abstract, p. 1: 'We propose a new simple network architecture, the Transformer, based solely on attention mechanisms'; §1 Introduction, p. 2), so the identification part of the attribution holds. However, nothing in our copy supports the substantive part of the claim as phrased: the paper reports no automated speech transcription, no word error rate, and no LibriSpeech/SwitchBoard results — its quality gains are BLEU on WMT 2014 English-to-German (28.4) and English-to-French (41.8) machine translation (§6.1, p. 8; Table 2, §5.4, p. 8) plus English constituency parsing F1 (§6.3, pp. 9-10). Searches for 'word error rate / speech recognition / transcription' and for parameter counts returned only translation and parsing passages; the largest model described (Transformer (big), 6 layers, d_model 1024, Table 3, §6.2, p. 8) is nowhere stated as a one-billion-parameter model, so that magnitude and the ASR/WER outcome are not attributable to this source. (LLM, conf=0.72)
- Not checked — no full text of the sourcecites [12] Graves, A. Sequence transduction with recurrent neural networks. Preprint at https://arxiv.org/abs/1211.3711 (2012). source
- Supportedcites [13] Graves, A., Mohamed, A.-R. & Hinton, G. Speech recognition with deep recurrent neural networks. In Proc. 2013 IEEE International Conference on Acoustics, Speech and Signal Processing 6645–6649 (IEEE, 2013) . source
The RNN Transducer section (p. 3) describes an RNN transducer for speech recognition, so the source supports the citation identifying that model type. The billion-parameter figure and the dataset-specific word-error-rate claim have separate citations in the citing sentence and are not attributed to this source. (LLM, conf=0.90)
- Partly supportedcites [16] Gulati, A. et al. Conformer: convolution-augmented transformer for speech recognition. Preprint at https://arxiv.org/abs/2005.08100 (2020). sourceExperimental — accuracy not yet measured
Our copy supports the WER half of the claim: Conformer, a convolution-augmented Transformer transducer for English ASR, is reported at state-of-the-art LibriSpeech WER of 2.1%/4.3% without LM and 1.9%/3.9% with an external LM (Abstract p. 1; §3.3 p. 3; Conclusion p. 4), beating the previously published Transformer Transducer. It does not support the 'up to one billion parameters' figure attached to this citation: §3.2 (p. 3) and Table 2 (p. 3) cap the models at 10.3M/30.7M/118.8M parameters, and the paper's thesis (§1 p. 1; Conclusion p. 4) is explicitly parameter efficiency — better accuracy with fewer parameters than prior work. The copy evaluates only LibriSpeech (no SwitchBoard), so any billion-scale or SwitchBoard element comes from other works, not this one. (LLM, conf=0.78)
- Supportedcites [2] Chan, W. et al. SpeechStew: simply mix all available speech recognition data to train one large neural network. Preprint at https://arxiv.org/abs/2104.02133 (2021). source
The citation is attached to the claim that the models contain up to one billion parameters. The Model section (p. 2) identifies SpeechStew as a Conformer RNN-T and says it uses a 1B-parameter configuration; the Experiments section (pp. 2–3) reports word error rates on LibriSpeech and Switchboard and describes its results as state-of-the-art or near state-of-the-art. (LLM, conf=0.90)
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Partly supported
The sources with a verdict support this claim; the rest weren’t checked yet, have no full text we could read, or their verdicts aren’t rated yet.
In the paper
“Analog-AI HW avoids these inefficiencies by leveraging arrays of non-volatile memory (NVM) to perform the ‘multiply and accumulate computation’ (MAC) operations which dominate these workloads directly in the memory3–7.”
§ An analog-AI chip for energy-efficient speech recognition and transcription > Main
What the sources say
- Not checked — no full text of the sourcecites [3] Ambrogio, S. et al. Equivalent-accuracy accelerated neural-network training using analogue memory. Nature 558, 60–67 (2018). source
- Supportedcites [4] Narayanan, P. et al. Fully on-chip MAC at 14 nm enabled by accurate row-wise programming of PCM-based weights and parallel vector-transport in duration-format. IEEE Trans. Electron. Devices 68, 6629–6636 (2021). source
The Introduction (p. 1) explains that resistive NVM arrays accelerate MAC operations by performing multiplication and accumulation where the data is stored, avoiding costly transfers between processing units and off-chip memory. The Conclusion (p. 7) reports a chip implementing in-memory MAC on analog PCM arrays. (LLM, conf=0.94)
- Supportedcites [5] Khaddam-Aljameh, R. et al. HERMES-core—a 1.59-TOPS/mm2 PCM on 14-nm CMOS in-memory compute core using 300-ps/LSB linearized CCO-based ADCs. IEEE J. Solid-State Circuits 57, 1027–1038 (2022). source
The Introduction (p. 1) says in-memory computing performs operations in the memory array, avoiding energy- and latency-costly data movement, and that matrix-vector multiplications dominate deep-neural-network inference. It explains that array rows supply inputs while columns sum partial products; the Introduction (p. 2) identifies nonvolatile PCM and ReRAM devices as suitable for these operations. (LLM, conf=0.93)
- Not checked — no full text of the sourcecites [6] Yao, P. et al. Fully hardware-implemented memristor convolutional neural network. Nature 577, 641–646 (2020). source
- Supportedcites [7] Wan, W. et al. A compute-in-memory chip based on resistive random-access memory. Nature 608, 504–512 (2022). source
The Abstract (page not supplied) says the chip stores AI-model weights in non-volatile RRAM and performs computation directly within it, avoiding power-hungry movement between separate memory and compute units. The Main section (page not supplied) identifies that in-memory computation as matrix-vector multiplication, which consists of multiply-and-accumulate operations. (LLM, conf=0.90)
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
A result the paper states — it cites no source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“We used a 14-nm analog inference chip to demonstrate SWeq end-to-end KWS on the Google Speech dataset using a fully analog set-up and a novel AB technique.”
§ An analog-AI chip for energy-efficient speech recognition and transcription > Conclusions
boosted- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“Each audio frame requires 2.4 μs in total, in the form of 8 time steps of 300 ns each (Fig. 3e); this is 7 times faster than the best-case latency currently reported by MLPerf8.”
§ An analog-AI chip for energy-efficient speech recognition and transcription > KWS task
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
48 more claims
Create a free account to see all 116 claims and to verify the 3 references still open.
- It may state two separate claims as one. 43
- Probably not a substantial claim. 24
- The quote only partly states this claim. 1
Citations
Cites
Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, L., & Polosukhin, I. (2017) · arXiv · doi:10.48550/arxiv.1706.03762 · Cited by 120,404
LeCun, Y., Bengio, Y., & Hinton, G. (2015) · Nature · doi:10.1038/nature14539 · Cited by 84,126
Neural Machine Translation by Jointly Learning to Align and Translate
Bahdanau, D., Cho, K., & Bengio, Y. (2014) · arXiv · doi:10.48550/arxiv.1409.0473 · Cited by 16,213
Speech recognition with deep recurrent neural networks
Graves, A., Mohamed, A.-R., & Hinton, G. (2013) · 2013 IEEE International Conference on Acoustics, Speech and Signal Processing · doi:10.1109/icassp.2013.6638947 · Cited by 8,893 · cited 1×
“In particular, transformer 1 and recurrent neural-network transducer (RNNT) 12,13,16 models containing up to one billion parameters 2 have produced a marked decrease in word error rate (WER) (and therefore much better accuracy) for the automated transcription of spoken English-language sentences,…”
Background“Librispeech: an ASR corpus based on public domain audio books,”
Panayotov, V., Chen, G., Povey, D., & Khudanpur, S. (2015) · 2015 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) · doi:10.1109/icassp.2015.7178964 · Cited by 6,072 · cited 5×
“We then implemented the MLPerf version of RNNT, a large data-center network, on Librispeech without any additional HWA retraining.”
MethodInfluentialGraves, A., Fernández, S., Gomez, F., & Schmidhuber, J. (2006) · Proceedings of the 23rd international conference on Machine learning - ICML '06 · doi:10.1145/1143844.1143891 · Cited by 5,370
Conformer: Convolution-augmented Transformer for Speech Recognition
Gulati, A., Qin, J., Chiu, C.-C., Parmar, N., Zhang, Y., Yu, J., Han, W., Wang, S., Zhang, Z., Wu, Y., et al. (2020) · arXiv · doi:10.48550/arxiv.2005.08100 · Cited by 3,150 · cited 2×
“In particular, transformer 1 and recurrent neural-network transducer (RNNT) 12,13,16 models containing up to one billion parameters 2 have produced a marked decrease in word error rate (WER) (and therefore much better accuracy) for the automated transcription of spoken English-language sentences,…”
BackgroundConformer: Convolution-augmented Transformer for Speech Recognition
Gulati, A., Qin, J., Chiu, C.-C., Parmar, N., Zhang, Y., Yu, J., Han, W., Wang, S., Zhang, Z., Wu, Y., et al. (2020) · Interspeech 2020 · doi:10.21437/interspeech.2020-3015 · Cited by 3,150 · cited 2×
“In particular, transformer 1 and recurrent neural-network transducer (RNNT) 12,13,16 models containing up to one billion parameters 2 have produced a marked decrease in word error rate (WER) (and therefore much better accuracy) for the automated transcription of spoken English-language sentences,…”
BackgroundContext-dependent pre-trained deep neural networks for large-vocabulary speech recognition
Dahl, G. E., Yu, D., Deng, L., & Acero, A. (2012) · IEEE Transactions on Audio Speech and Language Processing · doi:10.1109/tasl.2011.2134090 · Cited by 3,011 · cited 1×
“Furthermore, blocks 1(−1), 9(−9) and 2(−2), 10(−10) of Enc-LSTM0 Wx and Enc-LSTM1 Wx, and blocks 1(9), 17(25) (WP1(WP2)) and 2(10), 18(26) were summed in digital after on-chip analog MAC.”
MethodHuBERT: Self-Supervised Speech Representation Learning by Masked Prediction of Hidden Units
Hsu, W.-N., Bolte, B., Tsai, Y.-H. H., Lakhotia, K., Salakhutdinov, R., & Mohamed, A. (2021) · IEEE/ACM Transactions on Audio, Speech, and Language Processing · doi:10.1109/taslp.2021.3122291 · Cited by 2,983
Cited by
An integrated large-scale photonic accelerator with ultralow latency
Hua, S., Divita, E., Yu, S., Peng, B., Roques-Carmes, C., Su, Z., Chen, Z., Bai, Y., Zou, J., Zhu, Y., et al. (2025) · Nature · doi:10.1038/s41586-025-08786-6 · Cited by 199
“Second, like other technologies such as computing in memory 33 , photonic computing operates in the analogue domain, which means that it faces challenges in computation accuracy, especially for large, complex circuits.”
The growing memristor industry
Lanza, M., Pazos, S., Aguirre, F., Sebastian, A., Le Gallo, M., Alam, S. M., Ikegawa, S., Yang, J. J., Vianello, E., Chang, M.-F., et al. (2025) · Nature · doi:10.1038/s41586-025-08733-5 · Cited by 177
“2a) supporting larger networks (more than 1 million weights) have demonstrated promising inference accuracy results on more difficult benchmarks (such as image recognition on the CIFAR dataset) and high VMM energy efficiencies using RRAM 102–104 and PCM 91,105,106 .”
InfluentialMemristor-based hardware accelerators for artificial intelligence
Huang, Y., Ando, T., Sebastian, A., Chang, M., Yang, J. J., & Xia, Q. (2024) · Nature Reviews Electrical Engineering · doi:10.1038/s44287-024-00037-6 · Cited by 192
“Recent achievements in analogue in-memory computing chips have resulted in software-equivalent inference accuracies for industry-relevant AI use cases 30,31 and learning ability on the edge 32,84 .”
InfluentialNon-volatile 2D MoS2/black phosphorus heterojunction photodiodes in the near- to mid-infrared region
Zhu, Y., Wang, Y., Pang, X., Jiang, Y., Liu, X., Li, Q., Wang, Z., Liu, C., Hu, W., & Zhou, P. (2024) · Nature Communications · doi:10.1038/s41467-024-50353-6 · Cited by 150
E-waste Challenges of Generative Artificial Intelligence
Wang, P., Zhang, L.-Y., Tzachor, A., Masanet, E., & Chen, W.-Q. (2024) · doi:10.21203/rs.3.rs-3978528/v1 · Cited by 122
Metadata
- Authors
- Ambrogio, S., Narayanan, P., Okazaki, A., Fasoli, A., Mackin, C., Hosokawa, K., Nomura, A., Yasuda, T., Chen, A., Friz, A., Ishii, M., Luquin, J., Kohda, Y., Saulnier, N., Brew, K., Choi, S., Ok, I., Philip, T., Chan, V., Silvestre, C., Ahsan, I., Narayanan, V., Tsai, H., & Burr, G. W.
- Affiliations
- IBM Research - Almaden (US)
- IBM Research - Tokyo (JP)
- IBM Research - Albany (US)
- IBM Research - Thomas J. Watson Research Center (US)
- Corresponding author
- Ambrogio, S.
- ORCID
- Venue
- Nature
- Published
- 2023-08-23
- Volume
- 620
- Issue
- 7975
- Pages
- 768-775
- ISSN
- 0028-0836 (ISSN-L)
- 1476-4687
- PMC
- PMC10447234
- OpenAlex
- W4386100686
- Semantic Scholar
- b05e2df517662bffcd25a396deea07f21855fea3
- Cited by
- 269
- References
- 46
- Citations by year
- 2026: 85 · 2025: 105 · 2024: 71 · 2023: 8
- FWCI
- 19.32
- Type
- journal article
- Language
- en
- Open access
- hybrid
- doi.org
- Licence
- cc-by
- creativecommons.org/licenses/by/4.0 (tdm)
- creativecommons.org/licenses/by/4.0 (vor)
- Locations
- Nature (publishedVersion, cc-by)
- PubMed Central (publishedVersion, cc-by)
- PubMed (publishedVersion)
- RePEc: Research Papers in Economics (submittedVersion)
- Topics
- Physical Sciences › Engineering › Electrical and Electronic Engineering › Advanced Memory and Neural Computing
- Physical Sciences › Engineering › Electrical and Electronic Engineering › Low-power high-performance VLSI design
- Physical Sciences › Engineering › Electrical and Electronic Engineering › Analog and Mixed-Signal Circuit Design
- Keywords
- in-memory computing, phase-change memory, energy-efficient computing, speech recognition, speech transcription, RNN-T, keyword spotting, matrix-vector multiplication
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
The claim verdicts this report counts are experimental: their accuracy is not yet measured.
A language model drew the claims from the paper on 2026-10-10, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).
A second model labelled each claim on 2026-10-10 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).
Cited sources are checked with a language model’s help (gpt-6-sol and qwen3.8-omni-flash).
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
How to cite
Cite CoChat for the checks on this page, and the paper for its own findings.
- This page
- CoChat. (2026). Citation check: An analog-AI chip for energy-efficient speech recognition and transcription. https://app.cochat.ai/papers/10.1038/s41586-023-06337-5
- The paper
- Ambrogio, S., Narayanan, P., Okazaki, A., Fasoli, A., Mackin, C., Hosokawa, K., Nomura, A., Yasuda, T., Chen, A., Friz, A., Ishii, M., Luquin, J., Kohda, Y., Saulnier, N., Brew, K., Choi, S., Ok, I., Philip, T., Chan, V., . . . Burr, G. W. (2023). An analog-AI chip for energy-efficient speech recognition and transcription. Nature. https://doi.org/10.1038/s41586-023-06337-5
Data sources
Bibliographic data: Semantic Scholar, OpenAlex, Crossref.
Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).
Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.