PaperlayerView exact revision on arXiv

Source-linked AI summary

DLT-Corpus: A Large-Scale Text Collection for the Distributed Ledger Technology Domain

Walter Hernandez Cruz, Peter Devine, Nikhil Vadgama, Paolo Tasca, Jiahua Xu

arXiv:2602.22045v2cs.CL

TL;DR

Existing DLT NLP resources cover narrow tasks and omit substantial domain text. The paper constructs DLT-Corpus from scientific literature, USPTO patents, and social media, then analyzes technology diffusion and market–innovation relationships. Technologies generally appear first in scientific literature, while scientific and patent activity tracks overall market expansion more than short-term social-media sentiment.

  • Problem

    Existing DLT NLP datasets focus narrowly on tasks such as cryptocurrency price prediction and smart contracts, leaving scientific, patent, and technical resources underrepresented.

  • Method

    The paper aggregates 2.98 billion tokens from 22.12 million documents across scientific literature, USPTO patents, and social media, and analyzes diffusion and market–innovation relationships.

  • Results

    Technologies typically appear first in scientific literature before patents and social media, while scientific and patent activity tracks overall market expansion more than short-term sentiment.

  • Takeaways & Limitations

    DLT-Corpus provides a foundation for domain-specific NLP and integrated analysis of innovation emerging across research, commercial, and user communities.

  • Takeaways & Limitations

    The corpus focuses exclusively on English-language data from open-access scientific literature, patents, and social media.

Abstract

from arXiv · show

We introduce DLT-Corpus, the largest domain-specific text collection for Distributed Ledger Technology (DLT) research to date: 2.98 billion tokens from 22.12 million documents spanning scientific literature (37,440 publications), United States Patent and Trademark Office (USPTO) patents (49,023 filings), and social media (22 million posts). Existing Natural Language Processing (NLP) resources for DLT focus narrowly on cryptocurrency price prediction and smart contracts, leaving domain-specific language underexplored despite the sector's ~$3 trillion market capitalization and rapid technological evolution. We demonstrate DLT-Corpus' utility by analyzing patterns of technology emergence and market-innovation correlations. Findings reveal that technologies first appear in our scientific literature subset before reaching patents and social media, following traditional technology transfer patterns. While social media sentiment remains overwhelmingly bullish even during crypto winters, scientific and patent activity grows less tied to short-term sentiment, tracking overall market expansion in a virtuous cycle in which research precedes and enables economic growth that, in turn, funds further innovation. We release the DLT-Corpus and companion artifacts: LedgerBERT (+23% over BERT-base on DLT-specific Named Entity Recognition (NER) task), a sentiment analysis dataset of 23,301 crypto news headlines and descriptions, tools, and code.

1 Introduction

DLT-Corpus addresses fragmented, narrow DLT NLP resources with a large, multi-source domain corpus. Its analyses examine technology diffusion and links between innovation, market expansion, and community sentiment.

  • Motivation: Existing DLT datasets emphasize cryptocurrency price prediction, trading, fraud detection, sentiment analysis, and smart contracts while omitting substantial technical text.The overlooked sources include scientific publications, patents, technical documentation, and broader community discourse.
  • Contribution: DLT-Corpus contains 2.98 billion tokens from 22.12 million documents spanning scientific literature, USPTO patents, and social media.The corpus integrates technical specifications, economic mechanisms, community discourse, and governance frameworks.
  • Findings: DLT technologies and concepts typically appear first in scientific literature before reaching patents and social media, following traditional technology-transfer patterns.The paper highlights stablecoins, AMMs, and DEXs as examples of this sequence.
  • Artifacts: LedgerBERT improves over BERT-base by 23% on the DLT-specific NER task, alongside a 23,301-example sentiment-analysis dataset and released tools and code.LedgerBERT is developed through continued pre-training of SciBERT.
  • Contribution: The corpus supports integrated analysis of innovation diffusion across scientific publications, patents, and community discourse.Prior DLT analyses were fragmented across NER, patents, news, taxonomies, and systematic reviews.

4 Datasets

The datasets combine complementary DLT sources under accessibility and legal-compliance constraints, yielding broad coverage and strong domain specificity. They include large scientific, patent, social-media, and sentiment resources with structured metadata and documented processing.

  • Corpus composition: DLT-Corpus contains 2.98 billion tokens across 22.12 million documents from scientific literature, patents, and social media.The corpus includes 37,440 publications, 49,023 patent filings, and 22 million social-media posts.
  • Corpus composition: Academic and industry publications provide technical knowledge, patents reveal innovation trajectories, and social media capture community discourse and market dynamics.The multi-source design targets complementary aspects of the DLT ecosystem.
  • Scientific literature: The scientific subset retains 37,440 publications after domain filtering, duplicate removal, and manual review of marginally relevant documents.The subset spans 1978 to mid-2025.
  • Patents: The patents subset contains 49,023 USPTO documents spanning 1990 to mid-2025.The collection uses public USPTO records and focuses on US patents to facilitate research and commercial use.
  • Social media: The social-media subset contains 22,033,090 English-language posts collected before Twitter/X’s 2023 API restrictions.Posts span 2013 to mid-2023 and include timestamps and bullish, bearish, or neutral sentiment labels.
  • Corpus quality: DLT-Corpus has 8.7 times higher keyword density than general corpora and 98.7% document coverage for 361 DLT keywords.Its Jensen-Shannon divergence from general corpora ranges from 0.39 to 0.45.

5 Companion language model

The paper develops LedgerBERT through continued pre-training on DLT-Corpus and evaluates domain-specific NER alongside out-of-domain cryptocurrency-news sentiment. LedgerBERT improves DLT-specific NER while preserving sentiment performance relative to SciBERT.

  • Evaluation design: LedgerBERT is evaluated on in-domain DLT NER and out-of-domain cryptocurrency-news sentiment analysis.The two tasks test domain terminology learning and generalization beyond the corpus domains.
  • Training: Continued pre-training starts from SciBERT rather than training a model from scratch.SciBERT was selected as a multidisciplinary scientific starting point likely to contain some DLT-related material.
  • In-domain NER: 0.299 F1: LedgerBERT improves over BERT-base (0.243 F1) by 23% on strict DLT-specific NER.It also improves over SciBERT (0.289 F1) by 3.5% relative; strict matching requires exact entity boundaries and types.
  • Out-of-domain sentiment: LedgerBERT performs within 0.2% of SciBERT on market-direction sentiment, the primary out-of-domain metric.Performance is comparable across all sentiment dimensions, despite cryptocurrency news being absent from DLT-Corpus.
  • Interpretation: The combined NER and sentiment results are presented as domain specialization without catastrophic forgetting.The model gains DLT-specific knowledge while retaining general capabilities for out-of-domain tasks.

6 Dataset documentation

The dataset documentation follows standardized practices intended to support reproducibility and reuse. It applies the Datasheets for Datasets framework and FAIR principles.

  • Documentation: DLT-Corpus documentation follows the Datasheet for Datasets framework.The paper directs readers to additional documentation for DLT-Corpus and the sentiment analysis dataset.
  • FAIR alignment: The datasets are aligned with FAIR principles: Findable, Accessible, Interoperable, and Reusable.The stated audience includes human researchers and computational agents.
  • Documentation: Standardized documentation is intended to support reproducibility and future dataset use.The documentation covers the sentiment analysis dataset and DLT-Corpus through dedicated sections.

Motivation

The motivation is to address the lack of large-scale, domain-specific DLT text resources. The documented corpus combines scientific literature, patents, and social media while recording its scope and accessibility constraints.

  • Motivation: DLT-Corpus was created to address the lack of large-scale, domain-specific text corpora for DLT research and other research types.
  • Corpus scope: 2.98 billion tokens span three subsets: scientific literature, patents, and social media.The corpus content is further broken down into 37,440 scientific documents, 49,023 patent documents, and 22.03 million social-media documents.
  • Temporal coverage: The subsets cover scientific literature from 1978–2025, patents from 1990–2025, and social media from 2013 to mid-2023.
  • Scope boundaries: The corpus is English-only, with social-media posts after 2023 missing because of platform access restrictions.
  • Accessibility and privacy: All sources are publicly accessible, and social-media usernames are removed to protect privacy.The documentation states that the corpus contains no private or confidential data.

Collection

The collection aggregates scientific literature, USPTO patents, and social media using source-specific retrieval and preprocessing. The documented workflow includes relevance filtering, text standardization, privacy protection, and duplicate detection.

  • Scientific literature: Scientific literature is collected through Semantic Scholar API queries and filtered for domain relevance with a fine-tuned BERT model.Scientific documents are also converted from PDF to Markdown, language-filtered, and length-filtered.
  • Patents: Patents are retrieved from USPTO public databases using keyword searches.Patent text is extracted and formatting is standardized, including correction of encoding errors.
  • Social media: Social media is aggregated from previously published academic datasets and publicly available industry sources collected before Twitter/X’s 2023 API restrictions.Processing removes usernames, detects duplicates, and filters by language.
  • Dataset fields: The collection documentation includes dedicated field tables for scientific literature, patents, and social media.The cited tables are identified as Table 4, Table 5, and Table 6, respectively.

7 Analysis

The analyses examine how DLT technologies diffuse across scientific, patent, and social-media communities, and how document production and sentiment relate to cryptocurrency markets. Technologies consistently appear first in scientific literature, while research and patent activity is less tied to short-term sentiment than social-media activity.

  • Market-document correlations: Scientific literature, patents, and social media show strong positive correlations with annual cryptocurrency market capitalization: ρ=0.76, ρ=0.96, and ρ=0.98, respectively.The reported correlations use Spearman’s rank correlation across annual observations.
  • Market-document correlations: Scientific publications precede market expansion, with correlations weakening when the market leads and losing significance beyond two years.At three years, the market-leading correlation is ρ=0.47 with p>0.05.
  • Market-document correlations: Social media remains strongly correlated across nearly all lags, while patents show significant correlations whether patent activity leads or follows the market.Social-media correlation peaks concurrently at ρ=0.98; patent correlation peaks concurrently at ρ=0.97.
  • Technology diffusion across communities: Stablecoins, AMMs, and DEXs consistently appear first in scientific literature before reaching patents and social media.The pattern aligns with traditional technology transfer models in which research precedes commercial application and consumer adoption.
  • Cryptocurrency mentions vs. technology mentions: Bitcoin shows sustained user interest despite declining patents and plateauing publications, whereas Ethereum shows growing publications and patents alongside user interest.The contrast reflects Bitcoin’s mature consumer-asset trajectory and Ethereum’s continued innovation in smart contracts and DeFi.
  • Market sentiment and innovation activity: Even during crypto winters, social-media users remain overwhelmingly bullish, while patents and scientific publications follow trajectories largely independent of short-term sentiment.Innovation activity instead grows alongside overall market expansion.

8 Discussion

The discussion distinguishes user interest in cryptocurrencies from research and industry interest in underlying technologies. It interprets scientific priority, divergent cryptocurrency trajectories, and market-linked funding as a traditional technology-transfer pattern with a possible virtuous cycle.

  • Divergent community interests: New DLTs and concepts first appear in scientific literature before spreading to patents and the user community.The authors suggest recent scientific literature may help identify emerging technologies before they become mainstream.
  • Cryptocurrencies trajectory diverge: Bitcoin’s patent activity declines and scientific publications plateau despite sustained user interest, whereas Ethereum’s academic and patent activity grows alongside user interest.Ethereum’s pattern is associated with continued innovation through smart contracts and DeFi applications.
  • Cryptocurrencies trajectory diverge: XRP’s user engagement dropped sharply during legal challenges while research continued, and Hedera attracted academic interest with limited user engagement.The discussion presents these patterns as evidence that regulation, technological innovation, and market speculation independently shape the ecosystem.
  • Research creates economic value through a virtuous cycle: Foundational research precedes market expansion, while commercial innovation and community discourse respond more strongly to market conditions.The authors describe a cycle in which market growth may fund further industry research, patenting, adoption, and deployment.

9 Conclusions

The paper concludes that DLT-Corpus combines large-scale scientific, patent, and social-media text with analyses of technology diffusion and market relationships. Its findings support research-first technology transfer, while activity in science and patents tracks overall market expansion more than short-term sentiment.

  • Conclusions: DLT-Corpus contains 2.98 billion tokens from 22.12 million documents spanning scientific literature, USPTO patents, and social media.The collection includes academic publications and industry whitepapers within its scientific-literature component.
  • Conclusions: Technologies and concepts typically appear first in scientific literature before patents and social media, following traditional technology transfer patterns.Social-media sentiment remains overwhelmingly bullish during crypto winters, while scientific and patent activity is less tied to short-term sentiment.
  • Conclusions: The released artifacts include DLT-Corpus, a crowdsourced sentiment dataset, LedgerBERT, and reproducibility code.The artifacts support domain-specific NLP and innovation diffusion analysis for the DLT field.
  • Limitations: The corpus focuses exclusively on English-language data from open-access scientific literature, patents, and social media.The authors identify English as dominant for web content and nearly all scientific publications.
  • Limitations: Marginally relevant DLT papers may remain despite manual revision and removal of 570 papers from the filtered scientific-literature subset.This is presented as a domain-relevance filtering limitation.
  • Limitations: Prioritizing accessibility and legal compliance may limit subset sizes and collection from other data sources, including news.The trade-off is intended to reduce legal barriers for academic and commercial use.
  • Limitations: The sentiment dataset and DLT-Corpus could enable market manipulation or coordinated trading strategies if misused.The authors acknowledge this risk while emphasizing research transparency as the primary contribution.

A Additional datasets documentation

The supplementary documentation defines the intended research use, data scope, licensing, and limitations of the DLT resources. It also describes the sentiment dataset’s provenance, size, labels, and temporal and language coverage.

  • Resource scope: The resources are intended for NLP research, language-model development, innovation studies, and DLT text mining.The stated scope also includes social and computational linguistic studies.
  • Resource scope: The dataset is unsuitable for identifying individuals, creating investment advice without disclaimers, or applications requiring post-2023 social-media data.These boundaries reflect the documented intended-use constraints.
  • Resource scope: Misuse may enable market manipulation, so researchers should implement appropriate safeguards.The documentation identifies this as an impact risk.
  • Licensing and updates: The resources use mixed open-access scientific-literature licenses, USPTO patent terms, and CC-BY-NC 4.0 for social media.The social-media data was collected before Twitter/X terms changed in 2023.
  • Licensing and updates: The current release is a static snapshot; future versions may expand scientific literature and patents but likely will not include post-2023 social media.The update policy defines an important temporal boundary for reuse.
  • Sentiment analysis dataset: The sentiment dataset contains 23,301 examples with 1.85M tokens and three dimensions of crowdsourced labels.The dimensions cover market direction, content characteristics, and engagement quality, each with three categories.
  • Sentiment analysis dataset: The sentiment data covers English cryptocurrency news from January 2021 to May 2025 with no missing data or private content.Examples consist of publicly available headlines and brief descriptions voted on by CryptoPanic users.

Collection

The sentiment dataset uses crowdsourced cryptocurrency-user votes on CryptoPanic news, applies engagement and percentile-based quality controls, and supports DLT sentiment research. Its static, non-real-time snapshot is intended for research rather than unrestricted investment or voter-identification applications.

  • Data source and annotation: CryptoPanic news headlines and brief descriptions are annotated through crowdsourced voting by active cryptocurrency users across sentiment categories.Vote percentages are normalized by total engagement to incorporate domain-specific community input.
  • Label assignment: Articles below the 25th percentile are labeled negative, above the 75th percentile positive, and those between neutral for each sentiment dimension.Percentile boundaries are used to mitigate popularity bias.
  • Quality control: A minimum-vote threshold excludes articles with insufficient community engagement before annotation labels are used.This quality-control rule filters out low-engagement examples.
  • Intended use and access: The dataset is intended for sentiment analysis research, domain-specific model evaluation, and DLT market-sentiment studies.It is available through Hugging Face under a CC-BY-NC 4.0 research license.
  • Scope and limitations: The dataset is a static snapshot and is unsuitable for real-time sentiment applications, unrestricted investment systems, or identifying individual voters.Misuse may enable market manipulation, so safeguards and ethical guidelines are recommended.
Loading 2602.22045v2…