commoncrawl_citations_annotated_2022.csv

cc_project_author,post_title,cc_project_url,cc_project_category,post_date,keywords,abstract,cc_author_affiliation,cc_class,cc_snippet,cc_dataset_used,cc_derived_dataset_about,cc_derived_dataset_used,cc_derived_dataset_cited
"Vésteinn Snæbjarnarson, Haukur Barri Símonarson, Pétur Orri Ragnarsson, Svanhvít Lilja Ingólfsdóttir, Haukur Páll Jónsson, Vilhjálmur Þorsteinsson, Hafsteinn Einarsson – Miðeind ehf., Iceland; University of Iceland, Iceland",A Warm Start and a Clean Crawled Corpus -- A Recipe for Good Language Models,https://arxiv.org/abs/2201.05601,papers,20220101Z00:00:00,,"We train several language models for Icelandic, including IceBERT, that achieve state-of-the-art performance in a variety of downstream tasks, including part-of-speech tagging, named entity recognition, grammatical error detection and constituency parsing. To train the models we introduce a new corpus of Icelandic text, the Icelandic Common Crawl Corpus (IC3), a collection of high quality texts found online by targeting the Icelandic top-level-domain (TLD). Several other public data sources are also collected for a total of 16GB of Icelandic text. To enhance the evaluation of model performance and to raise the bar in baselines for Icelandic, we translate and adapt the WinoGrande dataset for co-reference resolution. Through these efforts we demonstrate that a properly cleaned crawled corpus is sufficient to achieve state-of-the-art results in NLP applications for low to medium resource languages, by comparison with models trained on a curated corpus. We further show that initializing models using existing multilingual models can lead to state-of-the-art results for some downstream tasks.","Miðeind ehf., Iceland; University of Iceland, Iceland","nlp/corpus-construction, nlp/language-model","3.1. The Icelandic Common Crawl Corpus¶ The Common Crawl Foundation is a non-profit organization that scrapes large semi-random subsets of the internet regularly and hosts timestamped and compressed dumps of the web online¹⁰ [¹⁰https://commoncrawl.org/the-data/get-started/]. Each dump contains billions of web pages occupying hundreds of terabytes. Parsing these files directly requires storage and computing power not directly available to most and can come at a significant financial cost. The foundation also hosts indices of URIs and their locations within the large zipped dump files. While these indices are also large, their processing is feasible with a few terabytes of storage.¶ 3.1.1. Extracting Icelandic Common Crawl data¶ The Common Crawl indices, which contain URI and byte offsets within the compressed dumps, are used to reduce the search space when looking for Icelandic texts. The Common Crawl Index Server has a public API¹¹ [¹¹https://index.commoncrawl.org/] where URIs can be queried based on attributes such as date, MIME-type and substring. Using the API eliminates the need to fetch the massive index files. To extract Icelandic, the .is pattern is targeted to match the Icelandic top level domain (TLD), resulting in 63.5 million retrieved pages with URIs and byte locations within the compressed Common Crawl dumps. The computational efficiency of our method can be attributed to these steps. Given the predominant use of the .is TLD for Icelandic web content, we assume that other TLDs have a much lower proportion of Icelandic content. That said, a nontrivial amount of text in Icelandic is still likely to be found outside the .is domain and could be extracted by, e.g., parsing the whole Common Crawl, albeit at a much higher computational cost.¶ By targeting only the byte-offsets corresponding to the Icelandic TLD we extract candidate websites that have a high proportion of Icelandic content. In total, the compressed content is 687GiB on disk. All dumps since the start of the Common Crawl in 2008 until March 2020 were included.¶ Plain text was extracted from the collected WARC (Web Archive format) files using jusText (Pomikálek, 2011)12 to remove boilerplate content and HTML tags.","CDX, WARC, ARC 2008 – March 2020",,,
"Mikel Artetxe, Itziar Aldabe, Rodrigo Agerri, Olatz Perez-de-Viñaspre, Aitor Soroa &ndash; Meta AI; HiTZ Center - Ixa, University of the Basque Country UPV/EHU",Does Corpus Quality Really Matter for Low-Resource Languages?,https://arxiv.org/abs/2203.08111,papers,20220101Z00:00:00,"Computation and Language (cs.CL), Artificial Intelligence (cs.AI), Machine Learning (cs.LG), FOS: Computer and information sciences, FOS: Computer and information sciences","The vast majority of non-English corpora are derived from automatically filtered versions of CommonCrawl. While prior work has identified major issues on the quality of these datasets (Kreutzer et al., 2021), it is not clear how this impacts downstream performance. Taking Basque as a case study, we explore tailored crawling (manually identifying and scraping websites with high-quality content) as an alternative to filtering CommonCrawl. Our new corpus, called EusCrawl, is similar in size to the Basque portion of popular multilingual corpora like CC100 and mC4, yet it has a much higher quality according to native annotators. For instance, 66% of documents are rated as high-quality for EusCrawl, in contrast with <33% for both mC4 and CC100. Nevertheless, we obtain similar results on downstream tasks regardless of the corpus used for pre-training. Our work suggests that NLU performance in low-resource languages is primarily constrained by the quantity rather than the quality of the data, prompting for methods to exploit more diverse data sources.","Meta AI; HiTZ Center - Ixa, University of the Basque Country UPV/EHU","nlp/corpus-construction, nlp/corpus-representativeness, nlp/corpus-quality, nlp/language-models, nlp/low-resource-languages","In this paper, we explore tailored crawling (i.e., manually identifying and scraping websites with high-quality content) as an alternative to filtering CommonCrawl. Taking Basque as a case study, we collect 12.5M documents from 33 websites with Creative Commons content. The resulting corpus, called EusCrawl, is similar in size to the Basque portion of CC100 and mC4, but it has substantially less issues and a higher perceived quality according to our blind audit with native annotators. However, we find that this improvement does not carry over to downstream tasks, as masked language models pre-trained on either corpora obtain similar results on 5 NLU benchmarks. Our results suggests that data quantity and domain play a more important role, prompting for methods to exploit more diverse sources of data in low-resource languages.",,,,
"Stella Biderman, Kieran Bicheno, Leo Gao &ndash; EleutherAI",Datasheet for the Pile,https://arxiv.org/abs/2201.07311,papers,20220101Z00:00:00,,"This datasheet describes the Pile, a 825 GiB dataset of human-authored text compiled by EleutherAI for use in large-scale language modeling. The Pile is comprised of 22 different text sources, ranging from original scrapes done for this project, to text data made available by the data owners, to third-party scrapes available online.",EleutherAI,"nlp/corpus-construction, nlp/corpus-datasheet, nlp/corpus-representativeness","Pile-CC: The Pile-CC dataset is a sample from the Common Crawl WARCs that has been converted to text using jusText [Endrédy and Novák, 2013].¶ [...] Pile-CC: The Pile-CC dataset was created to be included in the Pile. The underlying data comes from the Common Crawl, which was created to give people access to the wealth of information contained in the internet. Its creators were concerned that only data mining companies would be able to collect this data, and has the explicit aim of democratizing technology.¶ [...] Pile-CC: The data is sourced from Common Crawl, a non-profit 501(c)(3) organization founded by Gil Elbaz. The data from Common Crawl was processed by EleutherAI into Pile-CC.¶ [...] Pile-CC: Instances are webpages.¶ [...] Pile-CC: 54, 953, 117 documents, totaling 227.12 GiB.¶ [...] Pile-CC: A tiny fraction of the entire Common Crawl was included, chosen arbitrarily and heavily filtered as detailed in Gao et al. [2020].¶ [...] Pile-CC: Data in the Pile-CC dataset were scraped from websites by the Common Craw and then downloaded directly from the Common Craw by EleutherAI.¶ [...] Pile-CC: The earliest date of contents in Pile-CC is unknown.¶",,The-Pile-English,,
"Jonas Andersson Schwarz &ndash; Göteborgs Universitet, Sweden","The hitchhiker's guide Method handbook for quantification of online linguistic data in a country-specific context. Official research report, Linguistic Explorations of Societies (Work Package 1)",https://gupea.ub.gu.se/bitstream/handle/2077/70890/2022_1_Andersson%20Schwarz.pdf,papers,20220101Z00:00:00,,,"Göteborgs Universitet, Sweden","nlp/corpus-construction, nlp/corpus-representativeness","Central actors (in no particular order)¶ CommonCrawl. California-based non-profit organization that makes monthly crawls of the openly available Web and provides datasets and metadata to the public freely. The CommonCrawl corpus contains petabytes of data including raw web page data, metadata data and text data collected since 2011. Since 2012, CommonCrawl’s archive is hosted by Amazon Web Services as part of its Public Data Sets program. Every crawl contains around 300 terabytes of data and roughly 3 billion pages. In 2020, a filtered version of this CommonCrawl archive was used to train OpenAI’s GPT-3 language model.¶ [...] Similarly, CommonCrawl (2021) provides an aggregate listing the percentages of their database covered by each language – measured as the primary language of each html document, as identified by the Compact Language Detector 2 (CLD2) algorithm. This was included as a good benchmark to compare with.¶ [...] In comparison, when plotting the cur- rently stated language distribution of CommonCrawl (2021) in relation to the same population numbers of L1 and L2 speakers, the CommonCrawl distribution displays a similarly low kurtosis and skewness.",,,,
"Makoto Morishita, Katsuki Chousa, Jun Suzuki, Masaaki Nagata &ndash; NTT Communication Science Laboratories, NTT Corporation, Japan",JParaCrawl v3.0: A Large-scale English-Japanese Parallel Corpus,https://arxiv.org/abs/2202.12607,papers,20220101Z00:00:00,,"Most current machine translation models are mainly trained with parallel corpora, and their translation accuracy largely depends on the quality and quantity of the corpora. Although there are billions of parallel sentences for a few language pairs, effectively dealing with most language pairs is difficult due to a lack of publicly available parallel corpora. This paper creates a large parallel corpus for English-Japanese, a language pair for which only limited resources are available, compared to such resource-rich languages as English-German. It introduces a new web-based English-Japanese parallel corpus named JParaCrawl v3.0. Our new corpus contains more than 21 million unique parallel sentence pairs, which is more than twice as many as the previous JParaCrawl v2.0 corpus. Through experiments, we empirically show how our new corpus boosts the accuracy of machine translation models on various domains. The JParaCrawl v3.0 corpus will eventually be publicly available online for research purposes.","NTT Communication Science Laboratories, NTT Corporation, Japan","nlp/machine-translation, nlp/parallel-corpus, nlp/corpus-construction","Our method extracts parallel sentences from the web. Thus, the first step is finding a website that has parallel sentences. This method is based on the hypothesis that websites containing the same English and Japanese sentences might have parallel texts. To list such parallel websites, we analyzed all the Common Crawl text archive data released from March 2019 to August 2021³. [³During this period, the Common Crawl project released 25 archives, and their text size was about 212 TB.] We identified the language in the archive by CLD2⁴ [⁴ https://github.com/CLD2Owners/cld2] and listed 100,000 large websites that roughly have the same size of English and Japanese texts. For this step, we used extractor⁵ [⁵ 5https://github.com/paracrawl/extractor] that was provided by the ParaCrawl project.",,,,
"Imad LAKIM, Ebtesam Almazrouei, Ibrahim Abu Alhaol, Merouane Debbah, Julien Launay &ndash; TII, Abu Dhabi, Arabic Emirates; LightOn, Paris, France","A Holistic Assessment of the Carbon Footprint of Noor, a Very Large Arabic Language Model",https://openreview.net/forum?id=B-lS3zH8Zq,papers,20220101Z00:00:00,,"As ever larger language models grow more ubiquitous, it is crucial to consider their environmental impact. Characterised by extreme size and resource use, recent generations of models have been criticised for their voracious appetite for compute, and thus significant carbon footprint. Although reporting of carbon impact has grown more common in machine learning papers, this reporting is usually limited to compute resources used strictly for training. In this work, we propose a holistic assessment of the footprint of an extreme-scale language model, Noor. Noor is an ongoing project aiming to develop the largest multi-task Arabic language models--with up to 13B parameters--leveraging zero-shot generalisation to enable a wide range of downstream tasks via natural language instructions. We assess the total carbon bill of the entire project: starting with data collection and storage costs, including research and development budgets, pretraining costs, future serving estimates, and other exogenous costs necessary for this international cooperation. Notably, we find that inference costs and exogenous factors can have a significant impact on total budget. Finally, we discuss pathways to reduce the carbon footprint of extreme-scale models.","TII, Abu Dhabi, Arabic Emirates; LightOn, Paris, France","nlp/language-model, nlp/transformer-language-model, carbon-footprint","We use Common Crawl (CC) for acquiring large amounts of web data. Each CC dump is on average around 10TB, and we discard it immediately after processing it. On average, it takes 24 hours to fully process a dump: we used 21 dumps from CC, meaning we stored 210TB of data for 24hours, equivalent to 57 kWh of energy consumption. After processing the dumps, we got on average 1.2TB of data per dump, thus 25TB in total. Considering that this data will be stored for 6 months, we end up with 1.3 MWh of energy consumption for the bulk data. Note that we keep the processed data in all languages (not just Modern Standard Arabic).",,,,
"Asier Gutiérrez-Fandiño, David Pérez-Fernández, Jordi Armengol-Estapé, David Griol, Zoraida Callejas &ndash; LHF Labs; Universidad Autónoma de Madrid, Spain; University of Edinburgh, United Kingdom; Universidad de Granada, Spain",esCorpius: A Massive Spanish Crawling Corpus,https://ui.adsabs.harvard.edu/abs/2022arXiv220615147G,papers,20220101Z00:00:00,"Computer Science - Computation and Language, Computer Science - Artificial Intelligence",,"LHF Labs; Universidad Autónoma de Madrid, Spain; University of Edinburgh, United Kingdom; Universidad de Granada, Spain","nlp/corpus-construction, nlp/text-corpora","[…] In this paper, we introduce esCorpius, a Spanish crawling corpus obtained from near 1 Pb of Common Crawl data. It is the most extensive corpus in Spanish with this level of quality in the extraction, purification and deduplication of web textual content […] A total of 39,502 compressed WARC (Web Archive) from Common Crawl files were processed (see section 3.3 for more details). The compressed information occupied about 180 TB and the size of the processed decompressed information is estimated to be more than 0.8 PB. Prior to content deduplication, the downloaded corpus was composed of 106.768.594.753 words, 3.129.248.875 lines and 163.518.405 web pages. The deduplicated and cleaned corpus size is 346.262.072.705 bytes (322.5 GB), with 104.073.706 total number of lines, 50.040.055.322 tokens, 1.125.798.968 paragraphs and 2.421.598.201 sentences.",,,,
"Arnold Overwijk, Chenyan Xiong, Jamie Callan &ndash; Microsoft; Carnegie Mellon University",ClueWeb22: 10 Billion Web Documents with Rich Information,https://doi.org/10.1145/3477495.3536321,papers,20220101Z00:00:00,"clueweb, web corpus, dataset","ClueWeb22, the newest iteration of the ClueWeb line of datasets, is the result of more than a year of collaboration between industry and academia. Its design is influenced by the research needs of the academic community and the real-world needs of large-scale industry systems. Compared with earlier ClueWeb datasets, the ClueWeb22 corpus is larger, more varied, and has higher-quality documents. Its core is raw HTML, but it includes clean text versions of documents to lower the barrier to entry. Several aspects of ClueWeb22 are available to the research community for the first time at this scale, for example, visual representations of rendered web pages, parsed structured information from the HTML document, and the alignment of document distributions (domains, languages, and topics) to commercial web search.This talk shares the design and construction of ClueWeb22, and discusses its new features. We believe this newer, larger, and richer ClueWeb corpus will enable and support a broad range of research in IR, NLP, and deep learning.",Microsoft; Carnegie Mellon University,"cc-cited-not-used, nlp/corpus-construction, nlp/text-corpora, information-retrieval","One approach is to sift CommonCrawl data, eg, the C4 dataset used to pretrain T5 [10], which provides sufficient quantity, but the quality quickly becomes a concern. For example, the cleaned CommonCrawl reflects a quite weird distribution of the web [5]. Language models pretrained on C4 often perform worse than models pretrained on higher quality corpora at the same scale. With ClueWeb22, we aim to provide the web corpus for research in the near future. The design of ClueWeb22 emphasizes on these goals: 1) to reflect the distribution of the web in real scenarios; 2) to provide web pages at large quantity and also with high quality; 3) to enable new research directions by including information important in industry but previously not publicly available.",,,,
"Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, Todor Mihaylov, Myle Ott, Sam Shleifer, Kurt Shuster, Daniel Simig, Punit Singh Koura, Anjali Sridhar, Tianlu Wang, Luke Zettlemoyer &ndash; Meta AI",OPT: Open Pre-trained Transformer Language Models,https://arxiv.org/abs/2205.01068,papers,20220101Z00:00:00,,,Meta AI,"nlp/language-model, nlp/transformer-language-model, nlp/corpus-construction",,,,"CC-Stories, Pile-CC, CC-NEWS-RoBERTa-v2",
"Sylvain Lugeon, Tiziano Piccardi, Robert West &ndash; EPFL, Switzerland",Homepage2Vec: Language-Agnostic Website Embedding and Classification,https://ojs.aaai.org/index.php/ICWSM/article/download/19380/19152,papers,20220101Z00:00:00,,"Top-level domain. Some top-level domains (TLD) such as .edu or .biz can offer a good hint about the website's content. For example, a typical use case for .edu is university websites, whereas .biz is commonly associated with business activities. Following this intuition, we collected from Common Crawl,5 a large-scale sample of the Web, the 19 most frequent TLDs: .com, .org, .net, .info, .xyz, .club, .biz, .top, .edu, .online, .pro, .site, .vip, .icu, .buzz, .app, .asia, .gov, .space, excluding the country code TLD (ccTLD) because they indicate geographic origin, not website content. We represent this feature with a one-hot encoding vector of 19 dimensions.","EPFL, Switzerland","nlp/text-classification, web-site-classification",,,,,
"Johannes Zirngibl, Steffen Deusch, Patrick Sattler, Juliane Aulbach, Georg Carle, Mattijs Jonker &ndash; Technical University of Munich, Germany; University of Twente, The Netherlands","Domain Parking: Largely Present, Rarely Considered!",https://mediatum.ub.tum.de/1661842,papers,20220101Z00:00:00,,"Domain parking typically involves leveraging advertisements to generate revenue on otherwise inactive domain names. Their content is rarely of real value to users and tends to be highly similar across parked domains. They have commonalities beyond content alone: parked domains can share hosting and DNS infrastructure. Parking rarely receives special treatment in existing studies (e.g., content analyses or infrastructure concentration studies). While the presence and possible bias introduced by parked pages is sometimes acknowledged in studies, the studies still treat parked domains as any other, either because differentiation is infeasible, or because doing so is considered out-of-scope. We argue that the impact of parked domains on analyses regarding the current state and future development of the Internet should not be overlooked. In this paper, we motivate this argument through quantification, and take steps towards helping other researchers identify parked domains. We systematically collect a list of 82 parking services and develop DNS-based indicators to help identify parked domains. We next quantify the presence of parked domains, using large-scale DNS data containing hundreds of millions of registered domain names, representative for a significant part of the global DNS namespace. Overall, we pinpoint 60 M parked domains, which is a significant percentage of all names under consideration (23 %) and identify up to 4 % of domains from top lists to be parked. These findings demonstrate that the effect of parked pages is potentially pronounced. We also break down into the various parking services and DNS zones. This helps us demonstrate and further discuss the effect that domain parking can have on research and Internet consolidation.","Technical University of Munich, Germany; University of Twente, The Netherlands","web-science, internet/DNS, internet/domain-parking","Common Crawl While visual identification allowed us to validate the inferences to a reasonable extent, we wanted to upscale validation. Therefore, we consider Common Crawl (CC) data [21] [C. Crawl. (2022) The Common Crawl Corpus. [Online]. Available: https://commoncrawl.org/] and calculate the similarity of pages. Common Crawl is an open repository of web crawl data, collected at monthly intervals, accounting for hundreds of millions of unique domain names, and many more URLs. We consider CC data for Jan 2022 and the ∼60 M parked domains that we identify on Jan 28th, 2022. We extract the HTML content of parked pages from CC data, only considering URLs that contain exactly the registered domain. Furthermore, we require the crawl target to have been the landing page (i.e., the path of the URL is /) and also to have resulted in a useful response (i.e., HTTP status code of 200). Given these filters, ∼1.29 M HTML rich responses can be obtained. We extract visible text and tokenize it into words, remove stop words, apply lemmatization, and create a vector for the most-frequently used words for each page.",,,,
"Alexandra Sasha Luccioni, Frances Corry, Hamsini Sridharan, Mike Ananny, Jason Schultz, Kate Crawford &ndash; Hugging Face; University of Southern California, USA; New York University, USA; Microsoft Research, USA","A Framework for Deprecating Datasets: Standardizing Documentation, Identification, and Communication",https://doi.org/10.1145/3531146.3533086,papers,20220101Z00:00:00,"datasets, data stewardship data management dataset deprecation","Datasets are central to training machine learning (ML) models. The ML community has recently made significant improvements to data stewardship and documentation practices across the model development life cycle. However, the act of deprecating, or deleting, datasets has been largely overlooked, and there are currently no standardized approaches for structuring this stage of the dataset life cycle. In this paper, we study the practice of dataset deprecation in ML, identify several cases of datasets that continued to circulate despite having been deprecated, and describe the different technical, legal, ethical, and organizational issues raised by such continuations. We then propose a Dataset Deprecation Framework that includes considerations of risk, mitigation of impact, appeal mechanisms, timeline, post-deprecation protocols, and publication checks that can be adapted and implemented by the ML community. Finally, we propose creating a centralized, sustainable repository system for archiving datasets, tracking dataset modifications or deprecations, and facilitating practices of care and stewardship that can be integrated into research and publication processes.","Hugging Face; University of Southern California, USA; New York University, USA; Microsoft Research, USA","ai/ethics-of-machine-learning, nlp/text-corpora, nlp/corpus-construction, cc-cited-not-used","When it comes to filtering large text datasets scraped from the Web, given their sheer size (C4 represents 2.3 TB of data, whereas the Common Crawl has 139TB), filtering them is complex and time-consuming, although approaches have been proposed for reducing duplicates and train-test overlap [53]. [...] In practice, documenting and deprecating these datasets is akin to a game of whack-a-mole, since new versions of the Common Crawl come out every few months. Analyzing what they contain and their degrees of contamination through common evaluation tasks would take significant effort.",,,,
"Julia Kreutzer, Isaac Caswell, Lisa Wang, Ahsan Wahab, Daan van Esch, Nasanbayar Ulzii-Orshikh, Allahsera Tapo, Nishant Subramani, Artem Sokolov, Claytone Sikasote, Monang Setyawan, Supheakmungkol Sarin, Sokhar Samb, Benoît Sagot, Clara Rivera, Annette Rios, Isabel Papadimitriou, Salomey Osei, Pedro Ortiz Suarez, Iroro Orife, Kelechi Ogueji, Andre Niyongabo Rubungo, Toan Q. Nguyen, Mathias Müller, André Müller, Shamsuddeen Hassan Muhammad, Nanda Muhammad, Ayanda Mnyakeni, Jamshidbek Mirzakhalov, Tapiwanashe Matangira, Colin Leong, Nze Lawson, Sneha Kudugunta, Yacine Jernite, Mathias Jenny, Orhan Firat, Bonaventure F. P. Dossou, Sakhile Dlamini, Nisansa de Silva, Sakine Çabuk Ballı, Stella Biderman, Alessia Battisti, Ahmed Baruwa, Ankur Bapna, Pallavi Baljekar, Israel Abebe Azime, Ayodele Awokoya, Duygu Ataman, Orevaoghene Ahia, Oghenefego Ahia, Sweta Agrawal, Mofetoluwa Adeyemi &ndash; Google Research; Masakhane NLP; Turkic Interlingua; Haverford College; RobotsMali; Intel Labs; University of Zambia; Google; AIMS-AMMI; Inria; University of Zurich; Stanford University; Kwame Nkrumah University of Science and Technology; Sorbonne Université; Niger-Volta LTI; University of Waterloo; University of Electronic Science and Technology of China; University of Notre Dame; Bayero University Kano; University of South Florida; Hugging Face; Jacobs University Bremen; University of Moratuwa; EleutherAI; Obafemi Awolowo University; University of Ibadan; Instadeep; University of Maryland; Defence Space Administration Abuja",Quality at a Glance: An Audit of Web-Crawled Multilingual Datasets,https://doi.org/10.1162/tacl\_a\_00447,papers,20220101Z00:00:00,,"{With the success of large-scale pre-training and multilingual modeling in Natural Language Processing (NLP), recent years have seen a proliferation of large, Web-mined text datasets covering hundreds of languages. We manually audit the quality of 205 language-specific corpora released with five major public datasets (CCAligned, ParaCrawl, WikiMatrix, OSCAR, mC4). Lower-resource corpora have systematic issues: At least 15 corpora have no usable text, and a significant fraction contains less than 50\\% sentences of acceptable quality. In addition, many are mislabeled or use nonstandard/ambiguous language codes. We demonstrate that these issues are easy to detect even for non-proficient speakers, and supplement the human audit with automatic analyses. Finally, we recommend techniques to evaluate and improve multilingual corpora and discuss potential risks that come with low-quality data releases.}",Google Research; Masakhane NLP; Turkic Interlingua; Haverford College; RobotsMali; Intel Labs; University of Zambia; Google; AIMS-AMMI; Inria; University of Zurich; Stanford University; Kwame Nkrumah University of Science and Technology; Sorbonne Université; Niger-Volta LTI; University of Waterloo; University of Electronic Science and Technology of China; University of Notre Dame; Bayero University Kano; University of South Florida; Hugging Face; Jacobs University Bremen; University of Moratuwa; EleutherAI; Obafemi Awolowo University; University of Ibadan; Instadeep; University of Maryland; Defence Space Administration Abuja,"nlp/corpus-construction, nlp/web-as-corpus, nlp/parallel-corpus, nlp/low-resource-language","We selected the corpora for their multilinguality and the inclusion of understudied languages in NLP. With the exception of WikiMatrix and Paracrawl, all corpora are derived from CommonCrawl, and distinguish themselves by the choice of filtering methods, LangID and automatic alignment technology.",,"CCAligned-2020, Tensorflow-C4-Multilingual, OSCAR",,
"Julien Abadji, Pedro Ortiz Suarez, Laurent Romary, Benoît Sagot &ndash; Inria, France; Sorbonne Université, France",Towards a Cleaner Document-Oriented Multilingual Crawled Corpus,https://arxiv.org/abs/2201.06642,papers,20220101Z00:00:00,,"The need for raw large raw corpora has dramatically increased in recent years with the introduction of transfer learning and semi-supervised learning methods to Natural Language Processing. And while there have been some recent attempts to manually curate the amount of data necessary to train large language models, the main way to obtain this data is still through automatic web crawling. In this paper we take the existing multilingual web corpus OSCAR and its pipeline Ungoliant that extracts and classifies data from Common Crawl at the line level, and propose a set of improvements and automatic annotations in order to produce a new document-oriented version of OSCAR that could prove more suitable to pre-train large generative language models as well as hopefully other applications in Natural Language Processing and Digital Humanities.","Inria, France; Sorbonne Université, France","nlp/corpus-construction, nlp/web-as-corpus",,,OSCAR,,
"Wang Tongjing, Zhao Yin, Ziyu Bao, Evert Meijers &ndash; Utrecht University, The Netherlands; Delft University of Technology, The Netherlands",Dataset of intercity relationships between 293 Chinese cities extracted and classified on the basis of toponym co-occurrences on Common Crawl,https://www.researchgate.net/profile/Evert-Meijers/publication/362952059_Dataset_of_intercity_relationships_between_293_Chinese_cities_extracted_and_classified_on_the_basis_of_toponym_co-occurrences_on_Common_Crawl/links/6308bfc25eed5e4bd11f7938/Dataset-of-intercity-relationships-between-293-Chinese-cities-extracted-and-classified-on-the-basis-of-toponym-co-occurrences-on-Common-Crawl.pdf,papers,20220101Z00:00:00,"city networks, toponym co-occurrence, city relationship, geographical information retrieval","Although the importance of intercity relationships is theoretically acknowledged for cities’ socioeconomic development, the availability of such relational data often limits relevant urban studies. One of the new approaches of collecting city relational data is to extract the co-appearance of their place names from web texts. However, dealing with a gigantic web corpus is difficult for domain researchers given the complexities of processing terabytes of raw data. This paper develops an efficient and easy-to-follow method to extract a dataset of intercity relationships between 293 large Chinese cities applying the toponym co-occurrence method to a web archive. Our method successfully filters a 6.98 TB CC data set into a 202 GB single language text corpus. A highly-scalable Hadoop- based framework processes the full CC corpus utilizing a 1080 CPU cluster on the Amazon Elastic Map/Reduce infrastructure. To reveal more details of the intercity relationships, the intercity relationships are further classified into six categories: industry, information technology (IT), finance, research, culture, and government.","Utrecht University, The Netherlands; Delft University of Technology, The Netherlands","information retrieval, toponymy, dataset-creation",The data was retrieved from a Common Crawl raw corpus through a series of data processing. The web pages in this corpus that do not contain Chinese characteristics or Chinese placenames were filtered out based on keyword selection. The filtered Chinese corpus was 202 GB and the filtered Chinese corpus with placenames was about 139.5GB. Then we count the number of web pages where two city names co-appear. These intercity relationships were further classified into six categories using a lexicon-based classification method.,CC-MAIN-2019-18 (WET),,,
"Per E Kummervold, Freddy Wetjen, Javier de la Rosa &ndash; National Library of Norway (NLN), Norway",The Norwegian Colossal Corpus: A Text Corpus for Training Large Norwegian Language Models,http://www.lrec-conf.org/proceedings/lrec2022/pdf/2022.lrec-1.410.pdf,papers,20220101Z00:00:00,,,"National Library of Norway (NLN), Norway",nlp/corpus-construction,"Common Crawl (2022) is a non-profit organization that has been collecting data from the web and providing these archives to the public since 2011. Common Crawl-based datasets are popular for training transformer models and are the basis for the enormous 800GB Pile dataset (Gao, 2020), among others. There are extracted Norwegian datasets that are also based on Common Crawl. The Open Super-large Crawled Aggregated coRpus (OSCAR) (Suárez et al., 2019) contains 4.7GB (800M words) of Norwegian Bokmål and 54MB (9M words) of Norwegian Nynorsk. Using a cleaned version of Common Crawl, Google compiled a multilingual version of their English colossal corpus, called MC4 (2022), for training their mT5 model (Xue et al., 2020). The Norwegian part of that dataset is roughly 94GB (14B words). Both OSCAR and the MC4 datasets have been made available on Hugging Face (2022). Unfortunately, their respective licenses do not allow for redistribution within the NCC. To overcome this limitation, we are releasing scripts for the preparation, cleaning, deduplication, and formatting of these datasets, so they can be interleaved 3855with the NCC. By combining NCC with OSCAR and MC4, it should be possible to create a deduplicated Norwegian corpus with over 100GB of text (15B words).",,,OSCAR,
"Hanlin Li, Nicholas Vincent &ndash; Northwestern University, USA",Rethinking Data Governance: A Labor-Oriented Approach,https://criticalautomation.org/wp-content/uploads/2022/03/li-vincent-data-governance.pdf,papers,20220101Z00:00:00,,"The current data governance paradigm in which technology companies solely decide how user data is collected and used has introduced many issues to the tech sector. Prominent examples include information asymmetry about user data’s value, monopolistic practices enabled by data’s network effects, and power imbalance with respect to data aggregation and analysis. This work explicates how viewing users’ data-generating activities through a labor lens can help to mitigate these issues and provides corresponding design and research directions.","Northwestern University, USA","dataset-creation, data governance, user-generated content, artificial intelligence, machine learning, cc-cited-not-used","2.1 Information asymmetry about user data's value¶ The lack of transparency about user data's value helps make it possible for operators of for-profit computing systems to monetize user data and reap the bulk of its financial benefits. Currently, there exists a substantial gap between what data-driven technology companies know about user data's value and what users themselves do. For example, while social media platforms are well aware of the amount of financial benefits of user engagement, users do not have a window into how their collective attention and knowledge powers such businesses. This information asymmetry is further exacerbated by the fact that the vast majority of data that users produce during their interaction with modern technologies is rarely visible to themselves and is used downstream without their awareness and consent. For instance, the rise of AI technologies is possible largely due to the abundance of data unwittingly generated by the public for purposes other than enabling AI models. Prominent examples include Flickr photos [12], Wikipedia articles [14], and the Common Crawl dataset consisting of publicly available webpages [11]. In many of such cases, users produce data without being aware of its value and potential, giving technology companies the opportunity to extract an enormous amount of revenue from such data.",,,,
"Jiameng Pu, Zain Sarwar, Sifat Muhammad Abdullah, Abdullah Rehman, Yoonjin Kim, Parantapa Bhattacharya, Mobin Javed, Bimal Viswanath, Virginia Tech, LUMS Pakistan &ndash; Virginia Tech, USA; University of Chicago, USA; LUMS, Pakistan, University of Virginia, USA",Deepfake Text Detection: Limitations and Opportunities,https://jmpu.github.io/files/Deepfake%20Text%20Detection%20Limitations%20and%20Opportunities_CR.pdf,papers,20220101Z00:00:00,,,"Virginia Tech, USA; University of Chicago, USA; LUMS, Pakistan, University of Virginia, USA","nlp/text-classification, deep-fake-detection, misinformation, disinformation",,,,Grover-RealNews,
"Florian Hantke, Ben Stock &ndash; CISPA Helmholtz Center for Information Security, Germany",HTML Violations and Where to Find Them: A Longitudinal Analysis of Specification Violations in HTML,https://swag.cispa.saarland/papers/hantke2022violations.pdf,papers,20220101Z00:00:00,,,"CISPA Helmholtz Center for Information Security, Germany","web-science, internet-security","[...] we leveraged Common Crawl [22] to analyze more than 23K popular domains over the course of eight years. [...] the crawler framework first collects meta information for each of the listed domains using Common Crawl [22] as a basis for the following analyses (1). This Common Crawl approach makes it possible to take a look into the past and analyze old versions of websites as well as current snapshots. Unlike similar crawling studies before using the Internet Archive[32], with Common Crawl, we are not limited by rate limit issues as we can request the database and S3 bucket directly. This makes the process fast and enables to analyze nearly a thousand pages per minute from one IP address over multiple days. The meta information that the framework collects contains details on where an HTML document can be found in the Common Crawl’s dumps. For each domain, the framework collects meta information from up to 100 pages and hands them to the crawler.",,,,
"Todor Markov, Chong Zhang, Sandhini Agarwal, Tyna Eloundou, Teddy Lee, Steven Adler, Angela Jiang, Lilian Weng &ndash; OpenAI",A Holistic Approach to Undesired Content Detection in the Real World,https://arxiv.org/abs/2208.03274,papers,20220101Z00:00:00,,,OpenAI,"nlp/text-classification, nlp/corpus-construction, toxic content, hate speech",,,,,
"Joshua Reynolds, Adam Bates, Michael Bailey &ndash; New Mexico State University, USA; University of Illinois at Urbana-Champaign, USA; Georgia Institute of Technology, USA",Equivocal URLs: Understanding the Fragmented Space of URL Parser Implementations,https://link.springer.com/chapter/10.1007/978-3-031-17143-7_9,papers,20220101Z00:00:00,,,"New Mexico State University, USA; University of Illinois at Urbana-Champaign, USA; Georgia Institute of Technology, USA","computer-security/internet-security, web-security, URL parsing","We also surveyed ∼350 million URLs sampled uniformly and randomly from the approximately 3 billion URLs in Common Crawl's January 2022 URL Index [35]. [35 Kreymer, I., Chuang, G.: Announcing the common crawl index! (2015)]",,,,
"Mehmet Korkmaz, Emre Koçyiğit, Özgür Şahingöz, Banu Diri &ndash; Yildiz Technical University, Istanbul, Turkey; Biruni University, Istanbul, Turkey",A Hybrid Phishing Detection System Using Deep Learning-based URL and Content Analysis,https://www.eejournal.ktu.lt/index.php/elt/article/download/31197/15556,papers,20220101Z00:00:00,,,"Yildiz Technical University, Istanbul, Turkey; Biruni University, Istanbul, Turkey",computer-security/internet-security,,,,,
"Mohd Faizal Ab Razak, Mohd Izham Jaya, Ferda Ernawan, Ahmad Firdaus, Fajar Agung Nugroho &ndash; Universitas Dian Nuswantoro, Semarang, Indonesia",Comparative Analysis of Machine Learning Classifiers for Phishing Detection,https://ieeexplore.ieee.org/abstract/document/9930531/,papers,20220101Z00:00:00,,,"Universitas Dian Nuswantoro, Semarang, Indonesia",computer-security/internet-security,"… The source for this dataset is from the University Malaysia of Sarawak, compiled from PhishTank, OpenPhish, Alexa and Common Crawl. One method for detecting new phishing websites is to utilize heuristics such as the URL and CSS detection …",,,,
"L. Ranaldi, A. Nourbakhsh, F. Fallucchid, FM. Zanzotto &ndash; Guglielmo Marconi University, Roma, Italy; University of Rome Tor Vergata, Roma, Italy",C-OSINT: COVID-19 Open Source artificial INTelligence framework,https://ceur-ws.org/Vol-3260/paper16.pdf,papers,20220101Z00:00:00,,"With the emergence of COVID-19 disease worldwide, a market of the products related to this disease formed across the Internet. By the time these goods were in short supply, many uncontrolled Dark Web Marketplaces (DWM) were active in selling these products. At the same time, Dark Web Forums (DWF) became proxies for spreading false ideas, fake news about COVID-19, and advertising products sold in DWMs. This study investigates the activities entertained in the DWMs and DWFs to propose a learning-based model to distinguish them from their related counterparts on the surface web. To this end, we propose a COVID-19 Open Source artificial INTelligence framework (C-OSINT) to automatically collect and classify the activities done in DWMs and DWFs. Moreover, we corporate linguistic and stylistic solutions to leverage the classification performance between the content found in DWMs and DWFs and two surface web sources. Our results show that using syntactic and stylistic representation outperforms the Transformer based results over these domains.","Guglielmo Marconi University, Roma, Italy; University of Rome Tor Vergata, Roma, Italy",nlp/transformer-language-model; web-science/dark-web,,,,,
"Shuheng Liu, Alan Ritter &ndash; Georgia Institute of Technology",Do CoNLL-2003 Named Entity Taggers Still Work Well in 2023?,https://arxiv.org/abs/2212.09747,papers,20220101Z00:00:00,,,Georgia Institute of Technology,"nlp/named-entity-recognition, dataset-creation","Our dataset follows this distribution to collect Reuters news articles published between December 5th and 7th, 2020, collected from the Common Crawl Foundation³. [³http://commoncrawl.org/]",,,,
"Matyáš Boháček, Michal Bravanský, Filip Trhlík, Václav Moravec &ndash; Charles University, Prague, Czech Republic; Gymnasium of Johannes Kepler, Prague, Czech Republic; University College London, United Kingdom",Fine-grained Czech News Article Dataset: An Interdisciplinary Approach to Trustworthiness Analysis,https://arxiv.org/abs/2212.08550,papers,20220101Z00:00:00,,,"Charles University, Prague, Czech Republic; Gymnasium of Johannes Kepler, Prague, Czech Republic; University College London, United Kingdom","nlp/fake-news-detection, dataset-creation","Initially, we assembled a collection of almost 94, 000 articles by scraping URLs of 45 Czech news sources obtained from Common Crawl² [²https://commoncrawl.org/]. These sources included mainstream journalistic websites, tabloids, independent news outlets, and websites that are part of the disinformation ecosystem [ 26 ], capturing the full scope of journalistic content in the Czech Republic. [...] We applied multiple filters and balancing mechanisms to mitigate deficiencies caused by inherent flaws in Common Crawl, which reduced the dataset’s size from 94, 000 to 10, 000 items. This way, we also ensured that the data is as representative of the Czech news ecosystem and as diverse as possible.",,,,
"Mehtab Khan, Alex Hanna &ndash; Yale Law School, USA; Distributed AI Research Institute",The Subjects and Stages of AI Dataset Development: A Framework for Dataset Accountability,https://ssrn.com/abstract=4217148,papers,20220101Z00:00:00,,"There has been increased attention toward the datasets that are used to train and build AI technologies from the computer science and social science research communities, but less from legal scholarship. Both Large-Scale Language Datasets (LSLDs) and Large-Scale Computer Vision Datasets (LSCVDs) have been at the forefront of such discussions, due to recent controversies involving the use of facial recognition technologies, and the discussion of the use of publicly-available text for the training of massive models which generate human-like text. Many of these datasets serve as “benchmarks” to develop models that are used both in academic and industry research, while others are used solely for training models. The process of developing LSLDs and LSCVDs is complex and contextual, involving dozens of decisions about what kinds of data to collect, label, and train a model on, as well as how to make the data available to other researchers. However, little attention has been paid to mapping and consolidating the legal issues that arise at different stages of this process: when the data is being collected, after the data is used to build and evaluate models and applications, and how that data is distributed more widely. In this article, we offer four main contributions. First, we describe what kinds of objects these datasets are, how many different kinds exist, what types of modalities they encompass, and why they are important. Second, we provide more clarity about the stages of dataset development – a process that has thus far been subsumed within broader discussions about bias and discrimination – and the subjects who may be susceptible to harms at each point of development. Third, we provide a matrix of both the stages of dataset development and the subjects of dataset development, which traces the connections between stages and subjects. Fourth, we use this analysis to identify some basic legal issues that arise at the various stages in order to foster a better understanding of the dilemmas and tensions that arise at every stage. We situate our discussion within wider discussion of current debates and proposals related to algorithmic accountability. This paper fulfills an essential gap when it comes to comprehending the complicated landscape of legal issues connected to datasets and the gigantic AI models trained on them.","Yale Law School, USA; Distributed AI Research Institute","nlp/corpus-construction, dataset-creation, data-governance, privacy, legal/copyright",D. Common Crawl: Archiving the Whole Web The Common Crawl (CC) dataset is one of the most popular datasets used in the training of what have typically been called large language models. [...],,,,
"Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, Jenia Jitsev &ndash; LAION; UC Berkeley, USA; Gentec Data; TU Darmstadt, Germany; Hessian.AI; University of Washington, Seattle, USA; Technical University of Munich, Germany; Stability AI; EleutherAI; Juelich Supercomputing Center (JSC), Germany; Research Center Juelich (FZJ), Germany",LAION-5B: An open large-scale dataset for training next generation image-text models,https://arxiv.org/abs/2210.08402,papers,20220101Z00:00:00,,,"LAION; UC Berkeley, USA; Gentec Data; TU Darmstadt, Germany; Hessian.AI; University of Washington, Seattle, USA; Technical University of Munich, Germany; Stability AI; EleutherAI; Juelich Supercomputing Center (JSC), Germany; Research Center Juelich (FZJ), Germany","nlp/corpus-construction, nlp/multimodal-corpora","By starting from Common Crawl [1] and filtering this data source with an existing CLIP model, we derive a dataset consisting of three parts: 2.32 billion English image-text examples, 2.26 billion multilingual examples, and 1.27 billion examples that are not specific to a particular language (e.g., places, products, etc.). [...] To extract image-text pairs from Common Crawl, we parse the HTML IMG (image) tags from Common Crawl’s WAT metadata files.⁴ [⁴See https://commoncrawl.org/the-data/get-started/ for details of the metadata format.] Specifically, we focus on images with an alt-text so we can create image-text pair.",,LAION-5B,,
"{NLLB Team}, Marta R. Costa-jussà, James Cross, Onur Çelebi, Maha Elbayad, Kenneth Heafield, Kevin Heffernan, Elahe Kalbassi, Janice Lam, Daniel Licht, Jean Maillard, Anna Sun, Skyler Wang, Guillaume Wenzek, Al Youngblood, Bapi Akula, Loic Barrault, Gabriel Mejia Gonzalez, Prangthip Hansanti, John Hoffman, Semarley Jarrett, Kaushik Ram Sadagopan, Dirk Rowe, Shannon Spruit, Chau Tran, Pierre Andrews, Necip Fazil Ayan, Shruti Bhosale, Sergey Edunov, Angela Fan, Cynthia Gao, Vedanuj Goswami, Francisco Guzmán, Philipp Koehn, Alexandre Mourachko, Christophe Ropers, Safiyyah Saleem, Holger Schwenk, Jeff Wang &ndash; Meta AI; UC Berkeley, USA; Johns Hopkins University, USA",No Language Left Behind: Scaling Human-Centered Machine Translation,https://arxiv.org/abs/2207.04672,papers,20220101Z00:00:00,,"Driven by the goal of eradicating language barriers on a global scale, machine translation has solidified itself as a key focus of artificial intelligence research today. However, such efforts have coalesced around a small subset of languages, leaving behind the vast majority of mostly low-resource languages. What does it take to break the 200 language barrier while ensuring safe, high quality results, all while keeping ethical considerations in mind? In No Language Left Behind, we took on this challenge by first contextualizing the need for low-resource language translation support through exploratory interviews with native speakers. Then, we created datasets and models aimed at narrowing the performance gap between low and high-resource languages. More specifically, we developed a conditional compute model based on Sparsely Gated Mixture of Experts that is trained on data obtained with novel and effective data mining techniques tailored for low-resource languages. We propose multiple architectural and training improvements to counteract overfitting while training on thousands of tasks. Critically, we evaluated the performance of over 40,000 different translation directions using a human-translated benchmark, Flores-200, and combined human evaluation with a novel toxicity benchmark covering all languages in Flores-200 to assess translation safety. Our model achieves an improvement of 44% BLEU relative to the previous state-of-the-art, laying important groundwork towards realizing a universal translation system. Finally, we open source all contributions described in this work, accessible at https://github.com/facebookresearch/fairseq/tree/nllb.","Meta AI; UC Berkeley, USA; Johns Hopkins University, USA","nlp/corpus-construction, nlp/parallel-corpus, nlp/low-resource-language, nlp/language-identification","We begin with web data as our starting point, provided by CommonCrawl (CC)18 and ParaCrawl (Bañón et al., 2020).",,NLLB,,
"Shaden Smith, Mostofa Patwary, Brandon Norick, Patrick LeGresley, Samyam Rajbhandari, Jared Casper, Zhun Liu, Shrimai Prabhumoye, George Zerveas, Vijay Korthikanti, Elton Zhang, Rewon Child, Reza Yazdani Aminabadi, Julie Bernauer, Xia Song, Mohammad Shoeybi, Yuxiong He, Michael Houston, Saurabh Tiwary, Bryan Catanzaro &ndash; Microsoft; NVIDIA","Using DeepSpeed and Megatron to Train Megatron-Turing NLG 530B, A Large-Scale Generative Language Model",https://arxiv.org/abs/2201.11990,papers,20220101Z00:00:00,"Computation and Language (cs.CL), FOS: Computer and information sciences, FOS: Computer and information sciences",,Microsoft; NVIDIA,nlp/language-model,"Resources such as Common Crawl (CC) provide snapshots of the web which can be utilized as a source of language data. While these data sources contain an enormous amount of language data, they also require carefully designed preprocessing steps in order to select data which is of reasonable quality. As prior work has found (e.g., [9]), the quality of unfiltered Common Crawl data is lower than that of curated datasets and steps should be taken to increase the average quality of data selected from Common Crawl for LM pretraining. [...] Common Crawl: As mentioned previously, Common Crawl comprises an immense amount of data. We chose to process two snapshots, 2020-50 and 2021-04, with the aim of acquiring around 150B tokens of training data. The first step of this process is language detection [11] and text extraction from the raw HTML included in the Common Crawl WARC files¹. Following the rationale presented in [11], we used the pycld2² and jusText³ libraries for these tasks. [...] In addition to Common Crawl data, we leveraged a number of other previously generated datasets. From The Pile, we selected Books3, OpenWebText2, Stack Exchange, PubMed Abstracts, Wikipedia, Gutenberg (PG-19), BookCorpus2, NIH ExPorter, and Pile-CC datasets. We also included the CC-Stories and RealNews datasets used to train Megatron [63].",,,,
"Tom Alby, Robert Jäschke &ndash; Humboldt-Universität zu Berlin, Berlin, Germany",Analyzing the Web: Are Top Websites Lists a Good Choice for Research?,https://link.springer.com/chapter/10.1007/978-3-031-16802-4_2,papers,20220101Z00:00:00,,"The web has been a subject of research since its beginning, but it is difficult if not impossible to analyze the whole web, even if a database of all URLs would be freely accessible. Hundreds of studies have used commercial top websites lists as a shortcut, in particular the Alexa One Million Top Sites list. However, apart from the fact that Amazon decided to terminate Alexa, we question the usefulness of such lists for research as they have several shortcomings. Our analysis shows that top sites lists miss frequently visited websites and offer only little value for language-specific research. We present a heuristic-driven alternative based on the Common Crawl host-level web graph while also taking language-specific requirements into account.","Humboldt-Universität zu Berlin, Berlin, Germany","web-science, domain-ranking",,hyperlinkgraph/cc-main-2021-feb-apr-may/hostgraph,,,
"Olexandra Belz &ndash; Ivan Franko National University of Lviv, Ukraine",Use of schema.org micro-markup in e-commerce projects,http://baltijapublishing.lv/index.php/threeseas/article/view/1964/1973,papers,20220101Z00:00:00,,"The purpose of the article is to identify the most effective schema.org micro-markup schemes used in e-commerce projects. Methodology. The research included competitive intelligence among the leading online platforms operating in Europe in general and in Ukraine in particular. The study involved TOP-8 e-commerce projects in Ukraine and TOP-9 global cross-border marketplaces operating in Europe. The service validator.schema.org was chosen as the research tool. Results. The study showed that the most popular schema.org micro-markup format is JSON-LD. In general, 82.4% of the surveyed sites use JSON-LD microdata format. Some sites use two microdata formats: JSON-LD and Microdata. But none of the top online marketplaces use the RDFa micro-markup format. Popular marketplaces operating in Ukraine and Europe often use the same types of schema.org vocabulary. However, the frequency of using micro-markup by top marketplaces operating in Ukraine is much higher than the frequency of using micro-markup by top marketplaces operating in Europe. In addition, Ukrainian marketplaces use a much wider list of schema.org micro-markup properties than marketplaces operating in Europe. However, no online store has implemented the properties of advantages and disadvantages of goods recommended by Google in the scheme. Practical implications. The study suggests schema.org micro-markup schemes for homepage, category page, product page, about page, payment and delivery page, warranty and returns page, contact page and blog. The proposed templates of micro-markup schemes were validated using the validator.schema.org service. The study recommends using the JSON-LD format for semantic markup of website content. Value/originality. Implementation of effective semantic markup of site content will allow search engines to more accurately identify the information presented on the site. This, in turn, will improve the visibility of the online marketplace in the Search Engine Results Page of Google, Bing, Yahoo! etc.","Ivan Franko National University of Lviv, Ukraine","e-commerce, online marketplaces, linked data, schema.org annotations, SEO","Since 2008, the Common Crawl project has been crawling websites to collect web page data (extracting metadata and web page text). At the time of writing, the latest scan took place from November 26 to December 10, 2022. As a result of this scan, 3.35 billion web pages were processed and 420 petabytes of content were removed (Common Crawl, 2022). Both scientists and practitioners are working with the obtained data sets of the Common Crawl project.¶ On September 22, 2022, the Web Data Commons (WDC) project released the Schema.org Table Annotation Benchmark (SOTAB) for public download (Web Data Commons, 2022).",,,WebDataCommons,
"Minwoo Byeon, Beomhee Park, Haecheon Kim, Sungjun Lee, Woonhyuk Baek, Saehoon Kim &ndash; Kakao Brain, South Korea",Coyo-700m: Image-text pair dataset,https://github.com/kakaobrain/coyo-dataset,papers,20220101Z00:00:00,,We collected about 10 billion pairs of alt-text and image source in HTML documents in Common Crawl from Oct. 2020 to Aug. 2021. and eliminated uninformative pairs through the image and text level filtering process with minimal cost. The following figure outlines our data collection procedure.,"Kakao Brain, South Korea",nlp/multimodal-corpora,,"five CommonCrawl dumps, ranging from 2017 to 2020",COYO-700M,,
"Romal Thoppilan, Daniel De Freitas, Jamie Hall, Noam Shazeer, Apoorv Kulshreshtha, Heng-Tze Cheng, Alicia Jin, Taylor Bos, Leslie Baker, Yu Du, YaGuang Li, Hongrae Lee, Huaixiu Steven Zheng, Amin Ghafouri, Marcelo Menegali, Yanping Huang, Maxim Krikun, Dmitry Lepikhin, James Qin, Dehao Chen, Yuanzhong Xu, Zhifeng Chen, Adam Roberts, Maarten Bosma, Vincent Zhao, Yanqi Zhou, Chung-Ching Chang, Igor Krivokon, Will Rusch, Marc Pickett, Pranesh Srinivasan, Laichee Man, Kathleen Meier-Hellstern, Meredith Ringel Morris, Tulsee Doshi, Renelito Delos Santos, Toju Duke, Johnny Soraker, Ben Zevenbergen, Vinodkumar Prabhakaran, Mark Diaz, Ben Hutchinson, Kristen Olson, Alejandra Molina, Erin Hoffman-John, Josh Lee, Lora Aroyo, Ravi Rajakumar, Alena Butryna, Matthew Lamm, Viktoriya Kuzmina, Joe Fenton, Aaron Cohen, Rachel Bernstein, Ray Kurzweil, Blaise Aguera-Arcas, Claire Cui, Marian Croak, Ed Chi, Quoc Le &ndash; Google",LaMDA: Language Models for Dialog Applications,https://arxiv.org/abs/2201.08239,papers,20220101Z00:00:00,,,Google,"nlp/language-model, nlp/transformer-language-model","E Pre-training data composition¶ The pre-training data, called Infiniset, is a combination of dialog data from public dialog data and other public web documents. It consists of 2.97B documents and 1.12B dialogs with 13.39B utterances. The composition of the data is as follows: 50% dialogs data from public forums; 12.5% C4 data [11]; 12.5% code documents from sites related to programming like Q&A sites, tutorials, etc; 12.5% Wikipedia (English); 6.25% English web documents; and 6.25% Non-English web documents. The total number of words in the dataset is 1.56T. Note that this composition was chosen to achieve a more robust performance on dialog tasks (Section 4) while still keeping its ability to perform other tasks like code generation. As future work, we can study how the choice of this composition may affect the quality of some of the other NLP tasks performed by the model.",,,Tensorflow-C4,
"Mark Edward Phillips, Sawood Alam &ndash; University of North Texas, USA; Internet Archive, USA",Moving the End of Term Web Archive to the Cloud to Encourage Research Use and Reuse,https://digital.library.unt.edu/ark:/67531/metadc1998717/m2/1/high_res_d/EOT_WADL_2022.pdf,papers,20220101Z00:00:00,,"The End of Term Web (EOT) Archive is a collaborative project with a goal of collecting the United States federal web, loosely defined as .gov and .mil, every four years coinciding with presidential elections and often a transition in the Executive Branch of the government. In 2021 the End of Term team began to process the longitudinal web archive for EOT-2008, EOT-2012, EOT-2016, and EOT-2020 to move into the Amazon S3 storage service as part of the Amazon Open Data Program. This effort adopted tools, structures, and documentation developed by Common Crawl in an effort to maximize potential research access and reuse of existing tools and documentation. This paper presents the process of organizing, staging, processing, and moving these collections into the Amazon cloud.","University of North Texas, USA; Internet Archive, USA",web archive,,,,,
"Gilles Adda, Annelies Braffort, Ioana Vasilescu, François Yvon &ndash; Université Paris-Saclay, CNRS, LISN, Paris, France",Deliverable D1.14 Report on the French Language. European Language Equality (ELE); EU project no. LC- 01641480 – 101018166,https://european-language-equality.eu/wp-content/uploads/2022/03/ELE___Deliverable_D1_14__Language_Report_French_.pdf,papers,20220101Z00:00:00,,,"Université Paris-Saclay, CNRS, LISN, Paris, France","nlp/resources, French, nlp/language-models, nlp/text-corpora","The CommonCrawl project³⁷ [³⁷https://commoncrawl.org/] aggregates Web crawled data that is orders or magnitude larger than these resources for many languages; furthermore this corpus is being updated on a regular basis. By using parts of the French subset of CommonCrawl, possibly conjoined with the more curated corpora alluded to above has enabled to train large-scale BERT-style Language Models (LMs) – FlauBERT (Le et al., 2020) is built with a corpus containing about 12B running words, CamemBERT (Martin et al., 2020) uses the 22B words OSCAR, and these numbers continue to grow, albeit at a much slower pace than the corresponding English cor- pora.",,,,