With recent technological advancements, quantitative analysis has become an increasingly important area within professional sports. However, the manual process of collecting data on relevant match events like passes, goals and tacklings comes with considerable costs and limited consistency across providers, affecting both research and practice. In football, while automatic detection of events from positional data of the players and the ball could alleviate these issues, it is not entirely clear what accuracy current state-of-the-art methods realistically achieve because there is a lack of high-quality validations on realistic and diverse data sets. This paper adds context to existing research by validating a two-step rule-based pass and shot detection algorithm on four different data sets using a comprehensive validation routine that accounts for the temporal, hierarchical and imbalanced nature of the task. Our evaluation shows that pass and shot detection performance is highly dependent on the specifics of the data set. In accordance with previous studies, we achieve F-scores of up to 0.92 for passes, but only when there is an inherent dependency between event and positional data. We find a significantly lower accuracy with F-scores of 0.71 for passes and 0.65 for shots if event and positional data are independent. This result, together with a critical evaluation of existing methodologies, suggests that the accuracy of current football event detection algorithms operating on positional data is currently overestimated. Further analysis reveals that the temporal extraction of passes and shots from positional data poses the main challenge for rule-based approaches. Our results further indicate that the classification of plays into shots and passes is a relatively straightforward task, achieving F-scores between 0.83 to 0.91 ro rule-based classifiers and up to 0.95 for machine learning classifiers. We show that there exist simple classifiers that accurately differentiate shots from passes in different data sets using a low number of human-understandable rules. Operating on basic spatial features, our classifiers provide a simple, objective event definition that can be used as a foundation for more reliable event-based match analysis.

This dataset contains metadata (title, abstract, date of publication, field, etc) for around 1 million academic articles. Each record contains additional information on the country of study and whether the article makes use of data. Machine learning tools were used to classify the country of study and data use.

Our data source of academic articles is the Semantic Scholar Open Research Corpus (S2ORC) (Lo et al. 2020). The corpus contains more than 130 million English language academic papers across multiple disciplines. The papers included in the Semantic Scholar corpus are gathered directly from publishers, from open archives such as arXiv or PubMed, and crawled from the internet.

We placed some restrictions on the articles to make them usable and relevant for our purposes. First, only articles with an abstract and parsed PDF or latex file are included in the analysis. The full text of the abstract is necessary to classify the country of study and whether the article uses data. The parsed PDF and latex file are important for extracting important information like the date of publication and field of study. This restriction eliminated a large number of articles in the original corpus. Around 30 million articles remain after keeping only articles with a parsable (i.e., suitable for digital processing) PDF, and around 26% of those 30 million are eliminated when removing articles without an abstract. Second, only articles from the year 2000 to 2020 were considered. This restriction eliminated an additional 9% of the remaining articles. Finally, articles from the following fields of study were excluded, as we aim to focus on fields that are likely to use data produced by countries’ national statistical system: Biology, Chemistry, Engineering, Physics, Materials Science, Environmental Science, Geology, History, Philosophy, Math, Computer Science, and Art. Fields that are included are: Economics, Political Science, Business, Sociology, Medicine, and Psychology. This third restriction eliminated around 34% of the remaining articles. From an initial corpus of 136 million articles, this resulted in a final corpus of around 10 million articles.

Due to the intensive computer resources required, a set of 1,037,748 articles were randomly selected from the 10 million articles in our restricted corpus as a convenience sample.

The empirical approach employed in this project utilizes text mining with Natural Language Processing (NLP). The goal of NLP is to extract structured information from raw, unstructured text. In this project, NLP is used to extract the country of study and whether the paper makes use of data. We will discuss each of these in turn.

To determine the country or countries of study in each academic article, two approaches are employed based on information found in the title, abstract, or topic fields. The first approach uses regular expression searches based on the presence of ISO3166 country names. A defined set of country names is compiled, and the presence of these names is checked in the relevant fields. This approach is transparent, widely used in social science research, and easily extended to other languages. However, there is a potential for exclusion errors if a country’s name is spelled non-standardly.

The second approach is based on Named Entity Recognition (NER), which uses machine learning to identify objects from text, utilizing the spaCy Python library. The Named Entity Recognition algorithm splits text into named entities, and NER is used in this project to identify countries of study in the academic articles. SpaCy supports multiple languages and has been trained on multiple spellings of countries, overcoming some of the limitations of the regular expression approach. If a country is identified by either the regular expression search or NER, it is linked to the article. Note that one article can be linked to more than one country.

The second task is to classify whether the paper uses data. A supervised machine learning approach is employed, where 3500 publications were first randomly selected and manually labeled by human raters using the Mechanical Turk service (Paszke et al. 2019).[1] To make sure the human raters had a similar and appropriate definition of data in mind, they were given the following instructions before seeing their first paper:

Each of these documents is an academic article. The goal of this study is to measure whether a specific academic article is using data and from which country the data came.

There are two classification tasks in this exercise:

1. identifying whether an academic article is using data from any country

2. Identifying from which country that data came.

For task 1, we are looking specifically at the use of data. Data is any information that has been collected, observed, generated or created to produce research findings. As an example, a study that reports findings or analysis using a survey data, uses data. Some clues to indicate that a study does use data includes whether a survey or census is described, a statistical model estimated, or a table or means or summary statistics is reported.

After an article is classified as using data, please note the type of data used. The options are population or business census, survey data, administrative data, geospatial data, private sector data, and other data. If no data is used, then mark "Not applicable". In cases where multiple data types are used, please click multiple options.[2]

For task 2, we are looking at the country or countries that are studied in the article. In some cases, no country may be applicable. For instance, if the research is theoretical and has no specific country application. In some cases, the research article may involve multiple countries. In these cases, select all countries that are discussed in the paper.

We expect between 10 and 35 percent of all articles to use data.

The median amount of time that a worker spent on an article, measured as the time between when the article was accepted to be classified by the worker and when the classification was submitted was 25.4 minutes. If human raters were exclusively used rather than machine learning tools, then the corpus of 1,037,748 articles examined in this study would take around 50 years of human work time to review at a cost of $3,113,244, which assumes a cost of $3 per article as was paid to MTurk workers.

A model is next trained on the 3,500 labelled articles. We use a distilled version of the BERT (bidirectional Encoder Representations for transformers) model to encode raw text into a numeric format suitable for predictions (Devlin et al. (2018)). BERT is pre-trained on a large corpus comprising the Toronto Book Corpus and Wikipedia. The distilled version (DistilBERT) is a compressed model that is 60% the size of BERT and retains 97% of the language understanding capabilities and is 60% faster (Sanh, Debut, Chaumond, Wolf 2019). We use PyTorch to produce a model to classify articles based on the labeled data. Of the 3,500 articles that were hand coded by the MTurk workers, 900 are fed to the machine learning model. 900 articles were selected because of computational limitations in training the NLP model. A classification of “uses data” was assigned if the model predicted an article used data with at least 90% confidence.

The performance of the models classifying articles to countries and as using data or not can be compared to the classification by the human raters. We consider the human raters as giving us the ground truth. This may underestimate the model performance if the workers at times got the allocation wrong in a way that would not apply to the model. For instance, a human rater could mistake the Republic of Korea for the Democratic People’s Republic of Korea. If both humans and the model perform the same kind of errors, then the performance reported here will be overestimated.

The model was able to predict whether an article made use of data with 87% accuracy evaluated on the set of articles held out of the model training. The correlation between the number of articles written about each country using data estimated under the two approaches is given in the figure below. The number of articles represents an aggregate total of

Overview: 142: Areas used for sports, leisure and recreation purposes. Traceability (lineage): This dataset was produced with a machine learning framework with several input datasets, specified in detail in Witjes et al., 2022 (in review, preprint available at https://doi.org/10.21203/rs.3.rs-561383/v3 ) Scientific methodology: The single-class probability layers were generated with a spatiotemporal ensemble machine learning framework detailed in Witjes et al., 2022 (in review, preprint available at https://doi.org/10.21203/rs.3.rs-561383/v3 ). The single-class uncertainty layers were calculated by taking the standard deviation of the three single-class probabilities predicted by the three components of the ensemble. The HCL (hard class) layers represents the class with the highest probability as predicted by the ensemble. Usability: The HCL layers have a decreasing average accuracy (weighted F1-score) at each subsequent level in the CLC hierarchy. These metrics are 0.83 at level 1 (5 classes):, 0.63 at level 2 (14 classes), and 0.49 at level 3 (43 classes). This means that the hard-class maps are more reliable when aggregating classes to a higher level in the hierarchy (e.g. 'Discontinuous Urban Fabric' and 'Continuous Urban Fabric' to 'Urban Fabric'). Some single-class probabilities may more closely represent actual patterns for some classes that were overshadowed by unequal sample point distributions. Users are encouraged to set their own thresholds when postprocessing these datasets to optimize the accuracy for their specific use case. Uncertainty quantification: Uncertainty is quantified by taking the standard deviation of the probabilities predicted by the three components of the spatiotemporal ensemble model. Data validation approaches: The LULC classification was validated through spatial 5-fold cross-validation as detailed in the accompanying publication. Completeness: The dataset has chunks of empty predictions in regions with complex coast lines (e.g. the Zeeland province in the Netherlands and the Mar da Palha bay area in Portugal). These are artifacts that will be avoided in subsequent versions of the LULC product. Consistency: The accuracy of the predictions was compared per year and per 30km*30km tile across europe to derive temporal and spatial consistency by calculating the standard deviation. The standard deviation of annual weighted F1-score was 0.135, while the standard deviation of weighted F1-score per tile was 0.150. This means the dataset is more consistent through time than through space: Predictions are notably less accurate along the Mediterrranean coast. The accompanying publication contains additional information and visualisations. Positional accuracy: The raster layers have a resolution of 30m, identical to that of the Landsat data cube used as input features for the machine learning framework that predicted it. Temporal accuracy: The dataset contains predictions and uncertainty layers for each year between 2000 and 2019. Thematic accuracy: The maps reproduce the Corine Land Cover classification system, a hierarchical legend that consists of 5 classes at the highest level, 14 classes at the second level, and 44 classes at the third level. Class 523: Oceans was omitted due to computational constraints.

IntroductionFollowing the identification of Local Area Energy Planning (LAEP) use cases, this dataset lists the data sources and/or information that could help facilitate this research. View our dedicated page to find out how we derived this list: Local Area Energy Plan — UK Power Networks (opendatasoft.com)

Methodological Approach Data upload: a list of datasets and ancillary details are uploaded into a static Excel file before uploaded onto the Open Data Portal.

Quality Control Statement

Quality Control Measures include: Manual review and correct of data inconsistencies Use of additional verification steps to ensure accuracy in the methodology

Assurance Statement The Open Data Team and Local Net Zero Team worked together to ensure data accuracy and consistency.

Other Download dataset information: Metadata (JSON)

Definitions of key terms related to this dataset can be found in the Open Data Portal Glossary: https://ukpowernetworks.opendatasoft.com/pages/glossary/

Please note that "number of records" in the top left corner is higher than the number of datasets available as many datasets are indexed against multiple use cases leading to them being counted as multiple records.

Derived from over 150 years of lexical research, these comprehensive textual and audio data, focused on American English, provide linguistically annotated data. Ideal for NLP applications, LLM training and/or fine-tuning, as well as educational and game apps.

One of our flagship datasets, the American English data is expertly curated and linguistically annotated by professionals, with annual updates to ensure accuracy and relevance. The below datasets in American English are available for license:

1. American English Monolingual Dictionary Data
2. American English Synonyms and Antonyms Data
3. American English Pronunciations with Audio

Key Features (approximate numbers):

1. American English Monolingual Dictionary Data

Our American English Monolingual Dictionary Data is the foremost authority on American English, including detailed tagging and labelling covering parts of speech (POS), grammar, region, register, and subject, providing rich linguistic information. Additionally, all grammar and usage information is present to ensure relevance and accuracy.

• Headwords: 140,000
• Senses: 222,000
• Sentence examples: 140,000
• Format: XML and JSON format
• Delivery: Email (link-based file sharing) and REST API
• Updated frequency: annually

1. American English Synonyms and Antonyms Data

The American English Synonyms and Antonyms Dataset is a leading resource offering comprehensive, up-to-date coverage of word relationships in contemporary American English. It includes rich linguistic details such as precise definitions and part-of-speech (POS) tags, making it an essential asset for developing AI systems and language technologies that require deep semantic understanding.

• Synonyms: 600,000
• Antonyms: 22,000
• Format: XML and JSON format
• Delivery: Email (link-based file sharing) and REST API
• Updated frequency: annually

1. American English Pronunciations with Audio (word-level)

This dataset provides IPA transcriptions and clean audio data in contemporary American English. It includes syllabified transcriptions, variant spellings, POS tags, and pronunciation group identifiers. The audio files are supplied separately and linked where available for seamless integration - perfect for teams building TTS systems, ASR models, and pronunciation engines.

• Transcriptions (IPA): 250,000
• Audio files: 180,000
• Format: XLSX (for transcriptions), MP3 and WAV (audio files)
• Updated frequency: annually

Use Cases:

We consistently work with our clients on new use cases as language technology continues to evolve. These include NLP applications, TTS, dictionary display tools, games, translation machine, AI training and fine-tuning, word embedding, and word sense disambiguation (WSD).

If you have a specific use case in mind that isn't listed here, we’d be happy to explore it with you. Don’t hesitate to get in touch with us at Growth.OL@oup.com to start the conversation.

Pricing:

Oxford Languages offers flexible pricing based on use case and delivery format. Our datasets are licensed via term-based IP agreements and tiered pricing for API-delivered data. Whether you’re integrating into a product, training an LLM, or building custom NLP solutions, we tailor licensing to your specific needs.

Contact our team or email us at Growth.OL@oup.com to explore pricing options and discover how our language data can support your goals. Please note that some datasets may have rights restrictions. Contact us for more information.

About the sample:

To help you explore the structure and features of our dataset on this platform, we provide a sample in CSV and/or JSON formats for one of the presented datasets, for preview purposes only, as shown on this page. This sample offers a quick and accessible overview of the data's contents and organization.

Our full datasets are available in various formats, depending on the language and type of data you require. These may include XML, JSON, TXT, XLSX, CSV, WAV, MP3, and other file types. Please contact us (Growth.OL@oup.com) if you would like to receive the original sample with full details.

The National Land Cover Database (NLCD) is a land cover monitoring program providing land cover information for the United States. NLCD2016 extended temporal coverage to 15 years (2001–2016). We collected land cover reference data for the 2011 and 2016 nominal dates to report land cover accuracy for the NLCD2016 database 2011 and 2016 land cover components. We measured land cover accuracy at Level II and Level I, and change accuracy at Level I. For both the 2011 and 2016 land cover components, single-date Level II overall accuracies (OA) were 72% (standard error of ±0.9%) when agreement was defined as match between the map label and primary reference label only and 86% (± 0.7%) when agreement also included the alternate reference label. The corresponding level I OA for both dates were 79% (± 0.9%) and 91% (± 1.0%). For land cover change, the 2011–2016 user’s and producer’s accuracies (UA and PA) were ~ 75% for forest loss. PA for water loss, grassland loss, and grass gain were > 70% when agreement included a match between the map label and either the primary or alternate reference label. Depending on agreement definition and level of the classification hierarchy, OA for the 2011 land cover component of the NLCD2016 database was about 4% to 7% higher than OA for the 2011 land cover component of the NLCD2011 database, suggesting that the changes in mapping methodologies initiated for production of the NLCD2016 database have led to improved product quality.

Complete dataset of “Film Circulation on the International Film Festival Network and the Impact on Global Film Culture”

A peer-reviewed data paper for this dataset is in review to be published in NECSUS_European Journal of Media Studies - an open access journal aiming at enhancing data transparency and reusability, and will be available from https://necsus-ejms.org/ and https://mediarep.org

Please cite this when using the dataset.

Detailed description of the dataset:

1 Film Dataset: Festival Programs

The Film Dataset consists a data scheme image file, a codebook and two dataset tables in csv format.

The codebook (csv file “1_codebook_film-dataset_festival-program”) offers a detailed description of all variables within the Film Dataset. Along with the definition of variables it lists explanations for the units of measurement, data sources, coding and information on missing data.

The csv file “1_film-dataset_festival-program_long” comprises a dataset of all films and the festivals, festival sections, and the year of the festival edition that they were sampled from. The dataset is structured in the long format, i.e. the same film can appear in several rows when it appeared in more than one sample festival. However, films are identifiable via their unique ID.

The csv file “1_film-dataset_festival-program_wide” consists of the dataset listing only unique films (n=9,348). The dataset is in the wide format, i.e. each row corresponds to a unique film, identifiable via its unique ID. For easy analysis, and since the overlap is only six percent, in this dataset the variable sample festival (fest) corresponds to the first sample festival where the film appeared. For instance, if a film was first shown at Berlinale (in February) and then at Frameline (in June of the same year), the sample festival will list “Berlinale”. This file includes information on unique and IMDb IDs, the film title, production year, length, categorization in length, production countries, regional attribution, director names, genre attribution, the festival, festival section and festival edition the film was sampled from, and information whether there is festival run information available through the IMDb data.

2 Survey Dataset

The Survey Dataset consists of a data scheme image file, a codebook and two dataset tables in csv format.

The codebook “2_codebook_survey-dataset” includes coding information for both survey datasets. It lists the definition of the variables or survey questions (corresponding to Samoilova/Loist 2019), units of measurement, data source, variable type, range and coding, and information on missing data.

The csv file “2_survey-dataset_long-festivals_shared-consent” consists of a subset (n=161) of the original survey dataset (n=454), where respondents provided festival run data for films (n=206) and gave consent to share their data for research purposes. This dataset consists of the festival data in a long format, so that each row corresponds to the festival appearance of a film.

The csv file “2_survey-dataset_wide-no-festivals_shared-consent” consists of a subset (n=372) of the original dataset (n=454) of survey responses corresponding to sample films. It includes data only for those films for which respondents provided consent to share their data for research purposes. This dataset is shown in wide format of the survey data, i.e. information for each response corresponding to a film is listed in one row. This includes data on film IDs, film title, survey questions regarding completeness and availability of provided information, information on number of festival screenings, screening fees, budgets, marketing costs, market screenings, and distribution. As the file name suggests, no data on festival screenings is included in the wide format dataset.

3 IMDb & Scripts

The IMDb dataset consists of a data scheme image file, one codebook and eight datasets, all in csv format. It also includes the R scripts that we used for scraping and matching.

The codebook “3_codebook_imdb-dataset” includes information for all IMDb datasets. This includes ID information and their data source, coding and value ranges, and information on missing data.

The csv file “3_imdb-dataset_aka-titles_long” contains film title data in different languages scraped from IMDb in a long format, i.e. each row corresponds to a title in a given language.

The csv file “3_imdb-dataset_awards_long” contains film award data in a long format, i.e. each row corresponds to an award of a given film.

The csv file “3_imdb-dataset_companies_long” contains data on production and distribution companies of films. The dataset is in a long format, so that each row corresponds to a particular company of a particular film.

The csv file “3_imdb-dataset_crew_long” contains data on names and roles of crew members in a long format, i.e. each row corresponds to each crew member. The file also contains binary gender assigned to directors based on their first names using the GenderizeR application.

The csv file “3_imdb-dataset_festival-runs_long” contains festival run data scraped from IMDb in a long format, i.e. each row corresponds to the festival appearance of a given film. The dataset does not include each film screening, but the first screening of a film at a festival within a given year. The data includes festival runs up to 2019.

The csv file “3_imdb-dataset_general-info_wide” contains general information about films such as genre as defined by IMDb, languages in which a film was shown, ratings, and budget. The dataset is in wide format, so that each row corresponds to a unique film.

The csv file “3_imdb-dataset_release-info_long” contains data about non-festival release (e.g., theatrical, digital, tv, dvd/blueray). The dataset is in a long format, so that each row corresponds to a particular release of a particular film.

The csv file “3_imdb-dataset_websites_long” contains data on available websites (official websites, miscellaneous, photos, video clips). The dataset is in a long format, so that each row corresponds to a website of a particular film.

The dataset includes 8 text files containing the script for webscraping. They were written using the R-3.6.3 version for Windows.

The R script “r_1_unite_data” demonstrates the structure of the dataset, that we use in the following steps to identify, scrape, and match the film data.

The R script “r_2_scrape_matches” reads in the dataset with the film characteristics described in the “r_1_unite_data” and uses various R packages to create a search URL for each film from the core dataset on the IMDb website. The script attempts to match each film from the core dataset to IMDb records by first conducting an advanced search based on the movie title and year, and then potentially using an alternative title and a basic search if no matches are found in the advanced search. The script scrapes the title, release year, directors, running time, genre, and IMDb film URL from the first page of the suggested records from the IMDb website. The script then defines a loop that matches (including matching scores) each film in the core dataset with suggested films on the IMDb search page. Matching was done using data on directors, production year (+/- one year), and title, a fuzzy matching approach with two methods: “cosine” and “osa.” where the cosine similarity is used to match titles with a high degree of similarity, and the OSA algorithm is used to match titles that may have typos or minor variations.

The script “r_3_matching” creates a dataset with the matches for a manual check. Each pair of films (original film from the core dataset and the suggested match from the IMDb website was categorized in the following five categories: a) 100% match: perfect match on title, year, and director; b) likely good match; c) maybe match; d) unlikely match; and e) no match). The script also checks for possible doubles in the dataset and identifies them for a manual check.

The script “r_4_scraping_functions” creates a function for scraping the data from the identified matches (based on the scripts described above and manually checked). These functions are used for scraping the data in the next script.

The script “r_5a_extracting_info_sample” uses the function defined in the “r_4_scraping_functions”, in order to scrape the IMDb data for the identified matches. This script does that for the first 100 films, to check, if everything works. Scraping for the entire dataset took a few hours. Therefore, a test with a subsample of 100 films is advisable.

The script “r_5b_extracting_info_all” extracts the data for the entire dataset of the identified matches.

The script “r_5c_extracting_info_skipped” checks the films with missing data (where data was not scraped) and tried to extract data one more time to make sure that the errors were not caused by disruptions in the internet connection or other technical issues.

The script “r_check_logs” is used for troubleshooting and tracking the progress of all of the R scripts used. It gives information on the amount of missing values and errors.

4 Festival Library Dataset

The Festival Library Dataset consists of a data scheme image file, one codebook and one dataset, all in csv format.

The codebook (csv file “4_codebook_festival-library_dataset”) offers a detailed description of all variables within the Library Dataset. It lists the definition of variables, such as location and festival name, and festival categories, units of measurement, data sources and coding and missing data.

The csv file “4_festival-library_dataset_imdb-and-survey” contains data on all unique festivals collected from both IMDb and survey sources. This dataset appears in wide format, all information for each festival is listed in one row. This

This is the dataset split into three files. The data was collected from 527 organic food consumers. So, the files for factor analysis and logistic regression will have 527 respondents data while the file for clustering has only 401 organic food consumers data. Clustering was carried out only for those consuming organic food products.

The dataset shows the percentage of accuracy of traffic data collected from coils, for the year 2022. The 100% percentage means that in the reference time slot the coil has correctly detected the data. 0% means that the coil did not detect the data, for any reason, for the entire duration of the time slot. Intermediate percentages indicate partial detections within the reference time slot.The dataset must be used in parallel with the Vehicle detection through turns dataset - year 2022 where the traffic data are contained. To determine the accuracy of such traffic data, reference should be made to the combination of date and expiry code and the specific time slot in this dataset.

The SlangTrack (ST) Dataset is a novel, meticulously curated resource aimed at addressing the complexities of slang detection in natural language processing. This dataset uniquely emphasizes words that exhibit both slang and non-slang contexts, enabling a binary classification system to distinguish between these dual senses. By providing comprehensive examples for each usage, the dataset supports fine-grained linguistic and computational analysis, catering to both researchers and practitioners in NLP.

Key Features:

• Unique Words: 48,508
• Total Tokens: 310,170
• Average Post Length: 34.6 words
• Average Sentences per Post: 3.74

These features ensure a robust contextual framework for accurate slang detection and semantic analysis.

Significance of the Dataset:

1. Unified Annotation: The dataset offers consistent annotations across the corpus, achieving high Inter-Annotator Agreement (IAA) to ensure reliability and accuracy.
2. Addressing Limitations: It overcomes the constraints of previous corpora, which often lacked differentiation between slang and non-slang meanings or did not provide illustrative examples for each sense.
3. Comprehensive Coverage: Unlike earlier corpora that primarily supported dictionary-style entries or paraphrasing tasks, this dataset includes rich contextual examples from historical (COHA) and contemporary (Twitter) sources, along with multiple senses for each target word.
4. Focus on Dual Meanings: The dataset emphasizes words with at least one slang and one dominant non-slang sense, facilitating the exploration of nuanced linguistic patterns.
5. Applicability to Research: By covering both historical and modern contexts, the dataset provides a platform for exploring slang's semantic evolution and its impact on natural language processing.

Target Word Selection:

The target words were carefully chosen to align with the goals of fine-grained analysis. Each word in the dataset:

• It coexists in the slang SD wordlist and the Corpus of Historical American English (COHA).
• Has between 2 and 8 distinct senses, including both slang and non-slang meanings.
• Was cross-referenced using trusted resources such as:
  • Green's Dictionary of Slang
  • Urban Dictionary
  • Online Slang Dictionary
  • Oxford English Dictionary
• Features at least one slang and one dominant non-slang sense.
• Excludes proper nouns to maintain linguistic relevance and focus.

Data Sources and Collection:

1. Corpus of Historical American English (COHA):

• Historical examples were extracted from the cleaned version of COHA (CCOHA).
• Data spans the years 1980–2010, capturing the evolution of target words over time.

2. Twitter:

• Twitter was selected for its dynamic, real-time communication, offering rich examples of contemporary slang and informal language.
• For each target word, 1,000 examples were collected from tweets posted between 2010–2020, reflecting modern usage.

Dataset Scope:

The final dataset comprises ten target words, meeting strict selection criteria to ensure linguistic and computational relevance. Each word:

• Demonstrates semantic diversity, balancing slang and non-slang senses.
• Offers robust representation across both historical (COHA) and modern (Twitter) contexts.

The SlangTrack Dataset serves as a public resource, fostering research in slang detection, semantic evolution, and informal language processing. Combining historical and contemporary sources provides a comprehensive platform for exploring the nuances of slang in natural language.

Data Statistics:

The table below provides a breakdown of the total number of instances categorized as slang or non-slang for each target keyword in the SlangTrack (ST) Dataset.

Sample Texts from the Dataset:

The table below provides examples of sentences from the SlangTrack (ST) Dataset, showcasing both slang and non-slang usage of the target keywords. Each example highlights the context in which the target word is used and its corresponding category.

**Licenses**

The SlangTrack (ST) dataset is built using a combination of licensed and publicly available corpora. To ensure compliance with licensing agreements, all data has been extensively preprocessed, modified, and anonymized while preserving linguistic integrity. The dataset has been randomized and structured to support research in slang detection without violating the terms of the original sources.

The **original authors and data providers retain their respective rights**, where applicable. We encourage users to **review the licensing agreements** included with the dataset to understand any potential usage limitations. While some source corpora, such as **COHA, require a paid license and restrict redistribution**, our processed dataset is **legally shareable and publicly available** for **research and development purposes**.

The dataset has been collected in the frame of the Prac1 of the subject Tipology and Data Life Cycle of the Master's Degree in Data Science of the Universitat Oberta de Catalunya (UOC).

The dataset contains 25 variables and 52478 records corresponding to books on the GoodReads Best Books Ever list (the larges list on the site).

Original code used to retrieve the dataset can be found on github repository: github.com/scostap/goodreads_bbe_dataset

The data was retrieved in two sets, the first 30000 books and then the remainig 22478. Dates were not parsed and reformated on the second chunk so publishDate and firstPublishDate are representet in a mm/dd/yyyy format for the first 30000 records and Month Day Year for the rest.

Book cover images can be optionally downloaded from the url in the 'coverImg' field. Python code for doing so and an example can be found on the github repo.

The 25 fields of the dataset are:

| Attributes | Definition | Completeness |
| ------------- | ------------- | ------------- | 
| bookId | Book Identifier as in goodreads.com | 100 |
| title | Book title | 100 |
| series | Series Name | 45 |
| author | Book's Author | 100 |
| rating | Global goodreads rating | 100 |
| description | Book's description | 97 |
| language | Book's language | 93 |
| isbn | Book's ISBN | 92 |
| genres | Book's genres | 91 |
| characters | Main characters | 26 |
| bookFormat | Type of binding | 97 |
| edition | Type of edition (ex. Anniversary Edition) | 9 |
| pages | Number of pages | 96 |
| publisher | Editorial | 93 |
| publishDate | publication date | 98 |
| firstPublishDate | Publication date of first edition | 59 |
| awards | List of awards | 20 |
| numRatings | Number of total ratings | 100 |
| ratingsByStars | Number of ratings by stars | 97 |
| likedPercent | Derived field, percent of ratings over 2 starts (as in GoodReads) | 99 |
| setting | Story setting | 22 |
| coverImg | URL to cover image | 99 |
| bbeScore | Score in Best Books Ever list | 100 |
| bbeVotes | Number of votes in Best Books Ever list | 100 |
| price | Book's price (extracted from Iberlibro) | 73 |

MetaMath QA

Mathematical Questions for Large Language Models

By Huggingface Hub [source]

About this dataset

This dataset contains meta-mathematics questions and answers collected from the Mistral-7B question-answering system. The responses, types, and queries are all provided in order to help boost the performance of MetaMathQA while maintaining high accuracy. With its well-structured design, this dataset provides users with an efficient way to investigate various aspects of question answering models and further understand how they function. Whether you are a professional or beginner, this dataset is sure to offer invaluable insights into the development of more powerful QA systems!

More Datasets

For more datasets, click here.

Featured Notebooks

• 🚨 Your notebook can be here! 🚨!

How to use the dataset

Data Dictionary

The MetaMathQA dataset contains three columns: response, type, and query. - Response: the response to the query given by the question answering system. (String) - Type: the type of query provided as input to the system. (String) - Query:the question posed to the system for which a response is required. (String)

Preparing data for analysis

It’s important that before you dive into analysis, you first familiarize yourself with what kind data values are present in each column and also check if any preprocessing needs to be done on them such as removing unwanted characters or filling in missing values etc., so that it can be used without any issue while training or testing your model further down in your process flow.

##### Training Models using Mistral 7B

Mistral 7B is an open source framework designed for building machine learning models quickly and easily from tabular (csv) datasets such as those found in this dataset 'MetaMathQA ' . After collecting and preprocessing your dataset accordingly Mistral 7B provides with support for various Machine Learning algorithms like Support Vector Machines (SVM), Logistic Regression , Decision trees etc , allowing one to select from various popular libraries these offered algorithms with powerful overall hyperparameter optimization techniques so soon after selecting algorithm configuration its good practice that one use GridSearchCV & RandomSearchCV methods further tune both optimizations during model building stages . Post selection process one can then go ahead validate performances of selected models through metrics like accuracy score , F1 Metric , Precision Score & Recall Scores .

##### Testing phosphors :

After successful completion building phase right way would be robustly testing phosphors on different evaluation metrics mentioned above Model infusion stage helps here immediately make predictions based on earlier trained model OK auto back new test cases presented by domain experts could hey run quality assurance check again base score metrics mentioned above know asses confidence value post execution HHO updating baseline scores running experiments better preferred methodology AI workflows because Core advantage finally being have relevancy inexactness induced errors altogether impact low

Research Ideas

• Generating natural language processing (NLP) models to better identify patterns and connections between questions, answers, and types.
• Developing understandings on the efficiency of certain language features in producing successful question-answering results for different types of queries.
• Optimizing search algorithms that surface relevant answer results based on types of queries

Acknowledgements

If you use this dataset in your research, please credit the original authors. Data Source

License

License: CC0 1.0 Universal (CC0 1.0) - Public Domain Dedication No Copyright - You can copy, modify, distribute and perform the work, even for commercial purposes, all without asking permission. See Other Information.

Columns

File: train.csv | Column name | Description | |:--------------|:------------------------------------| | response | The response to the query. (String) | | type | The type of query. (String) |

Acknowledgements

If you use this dataset in your research, please credit the original authors. If you use this dataset in your research, please credit Huggingface Hub.

The LScDC (Leicester Scientific Dictionary-Core Dictionary)April 2020 by Neslihan Suzen, PhD student at the University of Leicester (ns433@leicester.ac.uk/suzenneslihan@hotmail.com)Supervised by Prof Alexander Gorban and Dr Evgeny Mirkes[Version 3] The third version of LScDC (Leicester Scientific Dictionary-Core) is formed using the updated LScD (Leicester Scientific Dictionary) - Version 3*. All steps applied to build the new version of core dictionary are the same as in Version 2** and can be found in description of Version 2 below. We did not repeat the explanation. The files provided with this description are also same as described as for LScDC Version 2. The numbers of words in the 3rd versions of LScD and LScDC are summarized below. # of wordsLScD (v3) 972,060LScDC (v3) 103,998 * Suzen, Neslihan (2019): LScD (Leicester Scientific Dictionary). figshare. Dataset. https://doi.org/10.25392/leicester.data.9746900.v3 ** Suzen, Neslihan (2019): LScDC (Leicester Scientific Dictionary-Core). figshare. Dataset. https://doi.org/10.25392/leicester.data.9896579.v2[Version 2] Getting StartedThis file describes a sorted and cleaned list of words from LScD (Leicester Scientific Dictionary), explains steps for sub-setting the LScD and basic statistics of words in the LSC (Leicester Scientific Corpus), to be found in [1, 2]. The LScDC (Leicester Scientific Dictionary-Core) is a list of words ordered by the number of documents containing the words, and is available in the CSV file published. There are 104,223 unique words (lemmas) in the LScDC. This dictionary is created to be used in future work on the quantification of the sense of research texts. The objective of sub-setting the LScD is to discard words which appear too rarely in the corpus. In text mining algorithms, usage of enormous number of text data brings the challenge to the performance and the accuracy of data mining applications. The performance and the accuracy of models are heavily depend on the type of words (such as stop words and content words) and the number of words in the corpus. Rare occurrence of words in a collection is not useful in discriminating texts in large corpora as rare words are likely to be non-informative signals (or noise) and redundant in the collection of texts. The selection of relevant words also holds out the possibility of more effective and faster operation of text mining algorithms.To build the LScDC, we decided the following process on LScD: removing words that appear in no more than 10 documents (

Title of data: CTE & SAE Data Inventory. Description of data: List of common data elements in clinical trials with domain, availability/completeness, occurrence in trials, semantic codes and definition. (XLSX 34 kb)

The ETCBC database of the Hebrew Bible (formerly known as WIVU database), contains the scholarly text of the Hebrew Bible with linguistic markup. A previous version can be found in EASY (see the link below). The present dataset is an improvement in many ways:

(A) it contains a new version of the data, called ETCBC4. The content has been heavily updated, with new linguistic annotations and a better organisation of them, and lots of additions and corrections as well.

(B) the data format is now Linguistic Annotation Framework (see below). This contrasts with the previous version, which has been archived as a database dump in a specialised format: Emdros (see the link below).

(C) a new tool, LAF-Fabric is added to process the ETCBC4 version directly from its LAF representation. The picture on this page shows a few samples what can be done with it.

(D) extensive documentation is provided, including a description of all the computing steps involved in getting the data in LAF format.

Since 2012 there is an ISO standard for the stand-off markup of language resources, Linguistic Annotation Framework (LAF).

As a result of the SHEBANQ project (see link below), funded by CLARIN-NL and carried out by the ETCBC and DANS, we have a created a tool, LAF-Fabric, by which we can convert EMDROS databases of the ETCBC into LAF and then do data analytic work by means of e.g. IPython notebooks. This has been used for the Hebrew Bible, but it can also be applied to the Syriac text in the CALAP (see link below).

This dataset contains a folder laf with the laf files, and the necessary declarations are contained in the folder decl. Among these declarations are feature declaration documents, in TEI format (see link below), with hyperlinks to concept definitions in ISOcat (see link below). For completeness, the ISOcat definitions are repeated in the feature declaration documents. These definitions are terse, and they are more fully documented in the folder documentation.

The USGS Protected Areas Database of the United States (PAD-US) is the nation's inventory of protected areas, including public open space and voluntarily provided, private protected areas, identified as an A-16 National Geospatial Data Asset in the Cadastral Theme (http://www.fgdc.gov/ngda-reports/NGDA_Datasets.html). PAD-US is an ongoing project with several published versions of a spatial database of areas dedicated to the preservation of biological diversity, and other natural, recreational or cultural uses, managed for these purposes through legal or other effective means. The geodatabase maps and describes public open space and other protected areas. Most areas are public lands owned in fee; however, long-term easements, leases, and agreements or administrative designations documented in agency management plans may be included. The PAD-US database strives to be a complete “best available” inventory of protected areas (lands and waters) including data provided by managing agencies and organizations. The dataset is built in collaboration with several partners and data providers (http://gapanalysis.usgs.gov/padus/stewards/). See Supplemental Information Section of this metadata record for more information on partnerships and links to major partner organizations. As this dataset is a compilation of many data sets; data completeness, accuracy, and scale may vary. Federal and state data are generally complete, while local government and private protected area coverage is about 50% complete, and depends on data management capacity in the state. For completeness estimates by state: http://www.protectedlands.net/partners. As the federal and state data are reasonably complete; focus is shifting to completing the inventory of local gov and voluntarily provided, private protected areas. The PAD-US geodatabase contains over twenty-five attributes and four feature classes to support data management, queries, web mapping services and analyses: Marine Protected Areas (MPA), Fee, Easements and Combined. The data contained in the MPA Feature class are provided directly by the National Oceanic and Atmospheric Administration (NOAA) Marine Protected Areas Center (MPA, http://marineprotectedareas.noaa.gov ) tracking the National Marine Protected Areas System. The Easements feature class contains data provided directly from the National Conservation Easement Database (NCED, http://conservationeasement.us ) The MPA and Easement feature classes contain some attributes unique to the sole source databases tracking them (e.g. Easement Holder Name from NCED, Protection Level from NOAA MPA Inventory). The "Combined" feature class integrates all fee, easement and MPA features as the best available national inventory of protected areas in the standard PAD-US framework. In addition to geographic boundaries, PAD-US describes the protection mechanism category (e.g. fee, easement, designation, other), owner and managing agency, designation type, unit name, area, public access and state name in a suite of standardized fields. An informative set of references (i.e. Aggregator Source, GIS Source, GIS Source Date) and "local" or source data fields provide a transparent link between standardized PAD-US fields and information from authoritative data sources. The areas in PAD-US are also assigned conservation measures that assess management intent to permanently protect biological diversity: the nationally relevant "GAP Status Code" and global "IUCN Category" standard. A wealth of attributes facilitates a wide variety of data analyses and creates a context for data to be used at local, regional, state, national and international scales. More information about specific updates and changes to this PAD-US version can be found in the Data Quality Information section of this metadata record as well as on the PAD-US website, http://gapanalysis.usgs.gov/padus/data/history/.) Due to the completeness and complexity of these data, it is highly recommended to review the Supplemental Information Section of the metadata record as well as the Data Use Constraints, to better understand data partnerships as well as see tips and ideas of appropriate uses of the data and how to parse out the data that you are looking for. For more information regarding the PAD-US dataset please visit, http://gapanalysis.usgs.gov/padus/. To find more data resources as well as view example analysis performed using PAD-US data visit, http://gapanalysis.usgs.gov/padus/resources/. The PAD-US dataset and data standard are compiled and maintained by the USGS Gap Analysis Program, http://gapanalysis.usgs.gov/ . For more information about data standards and how the data are aggregated please review the “Standards and Methods Manual for PAD-US,” http://gapanalysis.usgs.gov/padus/data/standards/ .

Additional file 1. List of files used for accuracy estimation of pointing correction. List of image files used for calculating parameters \(\bar{a}, \hat{a} {\text{ and }}\sigma\) a ÂŻ , a ^ and Ďƒ . This is also used for roughly estimating \(\theta_{1} {\text{ and }}\theta_{2}\) θ 1 and θ 2 . Names of the files displayed in this Additional file 1 may not be exactly identical to those stored in the publication sites, where the data files are released to the public. However, users can find the file uniquely because the observation date is included in each file name.

The National Land Cover Database (NLCD), a product suite produced through the Multi-resolution Land Characteristics (MRLC) consortium, is an operational land cover monitoring program. The release of NLCD2019 extends the database to 18 years. We collected land cover reference data for the 2016 and 2019 components of the NLCD2019 database at Level II and Level I of the classification hierarchy. For both dates, Level II land cover overall accuracies (OA) were 77.5% ± 1% (± value is the standard error) when agreement was defined as a match between the map label and primary reference label only and increased to 87.1% ± 0.7% when agreement was defined as a match between the map label and either the primary or alternate reference label. At Level I of the classification hierarchy, land cover OA was 83.1% ± 0.9% for both 2016 and 2019 when agreement was defined as a match between the map label and primary reference label only and increased to 90.3% ± 0.7% when agreement also included the alternate reference label. The Level II and Level I OA for the 2016 land cover in the NLCD2019 database were 5% higher compared to the 2016 land cover component of the NLCD2016 database when agreement was defined as a match between the map label and primary reference label only. No improvement was realized by the NLCD2019 database when agreement also included the alternate reference label. User’s accuracies (UA) for forest loss and grass gain were 70% when agreement included either the primary or alternate label, and UA was generally 50% for all other change themes. Producer’s accuracies (PA) were 70% for grass loss and gain and water gain and generally 50% for the other change themes.

CompanyKG is a heterogeneous graph consisting of 1,169,931 nodes and 50,815,503 undirected edges, with each node representing a real-world company and each edge signifying a relationship between the connected pair of companies.

Edges: We model 15 different inter-company relations as undirected edges, each of which corresponds to a unique edge type. These edge types capture various forms of similarity between connected company pairs. Associated with each edge of a certain type, we calculate a real-numbered weight as an approximation of the similarity level of that type. It is important to note that the constructed edges do not represent an exhaustive list of all possible edges due to incomplete information. Consequently, this leads to a sparse and occasionally skewed distribution of edges for individual relation/edge types. Such characteristics pose additional challenges for downstream learning tasks. Please refer to our paper for a detailed definition of edge types and weight calculations.

Nodes: The graph includes all companies connected by edges defined previously. Each node represents a company and is associated with a descriptive text, such as "Klarna is a fintech company that provides support for direct and post-purchase payments ...". To comply with privacy and confidentiality requirements, we encoded the text into numerical embeddings using four different pre-trained text embedding models: mSBERT (multilingual Sentence BERT), ADA2, SimCSE (fine-tuned on the raw company descriptions) and PAUSE.

Evaluation Tasks. The primary goal of CompanyKG is to develop algorithms and models for quantifying the similarity between pairs of companies. In order to evaluate the effectiveness of these methods, we have carefully curated three evaluation tasks:

Similarity Prediction (SP). To assess the accuracy of pairwise company similarity, we constructed the SP evaluation set comprising 3,219 pairs of companies that are labeled either as positive (similar, denoted by "1") or negative (dissimilar, denoted by "0"). Of these pairs, 1,522 are positive and 1,697 are negative.

Competitor Retrieval (CR). Each sample contains one target company and one of its direct competitors. It contains 76 distinct target companies, each of which has 5.3 competitors annotated in average. For a given target company A with N direct competitors in this CR evaluation set, we expect a competent method to retrieve all N competitors when searching for similar companies to A.

Similarity Ranking (SR) is designed to assess the ability of any method to rank candidate companies (numbered 0 and 1) based on their similarity to a query company. Paid human annotators, with backgrounds in engineering, science, and investment, were tasked with determining which candidate company is more similar to the query company. It resulted in an evaluation set comprising 1,856 rigorously labeled ranking questions. We retained 20% (368 samples) of this set as a validation set for model development.

Edge Prediction (EP) evaluates a model's ability to predict future or missing relationships between companies, providing forward-looking insights for investment professionals. The EP dataset, derived (and sampled) from new edges collected between April 6, 2023, and May 25, 2024, includes 40,000 samples, with edges not present in the pre-existing CompanyKG (a snapshot up until April 5, 2023).

Background and Motivation

In the investment industry, it is often essential to identify similar companies for a variety of purposes, such as market/competitor mapping and Mergers & Acquisitions (M&A). Identifying comparable companies is a critical task, as it can inform investment decisions, help identify potential synergies, and reveal areas for growth and improvement. The accurate quantification of inter-company similarity, also referred to as company similarity quantification, is the cornerstone to successfully executing such tasks. However, company similarity quantification is often a challenging and time-consuming process, given the vast amount of data available on each company, and the complex and diversified relationships among them.

While there is no universally agreed definition of company similarity, researchers and practitioners in PE industry have adopted various criteria to measure similarity, typically reflecting the companies' operations and relationships. These criteria can embody one or more dimensions such as industry sectors, employee profiles, keywords/tags, customers' review, financial performance, co-appearance in news, and so on. Investment professionals usually begin with a limited number of companies of interest (a.k.a. seed companies) and require an algorithmic approach to expand their search to a larger list of companies for potential investment.

In recent years, transformer-based Language Models (LMs) have become the preferred method for encoding textual company descriptions into vector-space embeddings. Then companies that are similar to the seed companies can be searched in the embedding space using distance metrics like cosine similarity. The rapid advancements in Large LMs (LLMs), such as GPT-3/4 and LLaMA, have significantly enhanced the performance of general-purpose conversational models. These models, such as ChatGPT, can be employed to answer questions related to similar company discovery and quantification in a Q&A format.

However, graph is still the most natural choice for representing and learning diverse company relations due to its ability to model complex relationships between a large number of entities. By representing companies as nodes and their relationships as edges, we can form a Knowledge Graph (KG). Utilizing this KG allows us to efficiently capture and analyze the network structure of the business landscape. Moreover, KG-based approaches allow us to leverage powerful tools from network science, graph theory, and graph-based machine learning, such as Graph Neural Networks (GNNs), to extract insights and patterns to facilitate similar company analysis. While there are various company datasets (mostly commercial/proprietary and non-relational) and graph datasets available (mostly for single link/node/graph-level predictions), there is a scarcity of datasets and benchmarks that combine both to create a large-scale KG dataset expressing rich pairwise company relations.

Source Code and Tutorial:https://github.com/llcresearch/CompanyKG2

Paper: to be published

Derived from over 150 years of lexical research, these comprehensive textual and audio data, focused on British English, provide linguistically annotated data. Ideal for NLP applications, Machine Learning (ML), LLM training and/or fine-tuning, as well as educational and game apps.

Our British English language datasets are meticulously curated and annotated by experienced linguists and language experts, ensuring exceptional accuracy, consistency, and linguistic depth. The below datasets in British English are available for license:

1. British English Monolingual Dictionary Data
2. British English Synonyms and Antonyms Data
3. British English Pronunciations with Audio

Key Features (approximate numbers):

1. British English Monolingual Dictionary Data

Our British English monolingual dataset delivers clear, reliable definitions and authentic usage examples, featuring a high volume of headwords and in-depth coverage of the British English variant of English. As one of the world’s most authoritative lexical resources, it’s trusted by leading academic, AI, and language technology organizations.

• Headwords: 146,000
• Senses: 230,000
• Sentence examples: 149,000
• Format: XML and JSON format
• Delivery: Email (link-based file sharing) and REST API
• Updated frequency: twice a year

1. British English Synonyms and Antonyms Data

This British English language dataset offers a rich collection of synonyms and antonyms, accompanied by detailed definitions and part-of-speech (POS) annotations, making it a comprehensive resource for NLP tasks such as semantic search, word sense disambiguation, and language generation.

• Synonyms: 600,000
• Antonyms: 22,000
• Usage Examples: 39,000
• Format: XML and JSON format
• Delivery: Email (link-based file sharing)
• Updated frequency: annually

1. British English Pronunciations with audio (word-level)

This dataset provides IPA transcriptions and clean audio data in contemporary British English. It includes syllabified transcriptions, variant spellings, POS tags, and pronunciation group identifiers. The audio files are supplied separately and linked where available for seamless integration - perfect for teams building TTS systems, ASR models, and pronunciation engines.

• Transcriptions (IPA): 250,000
• Audio files: 180,000
• Format: XLSX (for transcriptions), MP3 and WAV (audio files)
• Updated frequency: annually

Use Cases:

We consistently work with our clients on new use cases as language technology continues to evolve. These include NLP applications, TTS, dictionary display tools, games, translation machine, AI training and fine-tuning, word embedding, and word sense disambiguation (WSD).  

If you have a specific use case in mind that isn't listed here, we’d be happy to explore it with you. Don’t hesitate to get in touch with us at Growth.OL@oup.com to start the conversation.

Pricing:  

Oxford Languages offers flexible pricing based on use case and delivery format. Our datasets are licensed via term-based IP agreements and tiered pricing for API-delivered data. Whether you’re integrating into a product, training an LLM, or building custom NLP solutions, we tailor licensing to your specific needs. 

Contact our team or email us at Growth.OL@oup.com to explore pricing options and discover how our language data can support your goals. 

Please note that some datasets may have rights restrictions. Contact us for more information. 

About the sample: 

To help you explore the structure and features of our dataset on this platform, we provide a sample in CSV and/or JSON formats for one of the presented datasets, for preview purposes only, as shown on this page. This sample offers a quick and accessible overview of the data's contents and organization.  

Our full datasets are available in various formats, depending on the language and type of data you require. These may include XML, JSON, TXT, XLSX, CSV, WAV, MP3, and other file types. Please contact us (Growth.OL@oup.com) if you would like to receive the original sample with full details.