Slovenian datasets for contextual synonym and antonym detection can be used for training machine learning classifiers as described in the MSc thesis of Jasmina Pegan "Semantic detection of synonyms and antonyms with contextual embeddings" (https://repozitorij.uni-lj.si/IzpisGradiva.php?id=141456). Datasets contain example pairs of synonyms and antonyms in contexts together with additional information on a sense pair. Candidates for synonyms and antonyms were retrieved from the dataset created in the BSc thesis of Jasmina Pegan "Antonym detection with word embeddings" (https://repozitorij.uni-lj.si/IzpisGradiva.php?id=110533). Example sentences were retrieved from The comprehensive Slovenian-Hungarian dictionary (VSMS) (https://www.clarin.si/repository/xmlui/handle/11356/1453). Each dataset is class balanced and contains an equal amount of examples and counterexamples. An example is a pair of example sentences where the two words are synonyms/antonyms. A counterexample is a pair of example sentences where two words are not synonyms/antonyms. Note that a word pair can be synonymous or antonymous in some sense of the two words (but not in the given context).

Datasets are divided into two categories, datasets for synonyms and datasets for antonyms. Each category is further divided into base and updated datasets. These contain three dataset files: train, validation and test dataset. Base datasets include only manually-reviewed sense pairs. These are generated from all pairs of VSMS sense examples for all confirmed pairs of antonym and synonym senses. Updated datasets include automatically generated sense pairs while constraining the maximal number of examples per word. In this way, the dataset is more balanced word-wise, but is not fully manually-reviewed and contains less accurate data.

A single dataset entry contains the information on the base word, followed by data on synonym/antonym candidate. The last column discerns whether the sense pair is a pair of synonyms/antonyms or not. More details on this can be found inside the included README file.

LScDC Word-Category RIG MatrixApril 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 MirkesGetting StartedThis file describes the Word-Category RIG Matrix for theLeicester Scientific Corpus (LSC) [1], the procedure to build the matrix and introduces the Leicester Scientific Thesaurus (LScT) with the construction process. The Word-Category RIG Matrix is a 103,998 by 252 matrix, where rows correspond to words of Leicester Scientific Dictionary-Core (LScDC) [2] and columns correspond to 252 Web of Science (WoS) categories [3, 4, 5]. Each entry in the matrix corresponds to a pair (category,word). Its value for the pair shows the Relative Information Gain (RIG) on the belonging of a text from the LSC to the category from observing the word in this text. The CSV file of Word-Category RIG Matrix in the published archive is presented with two additional columns of the sum of RIGs in categories and the maximum of RIGs over categories (last two columns of the matrix). So, the file ‘Word-Category RIG Matrix.csv’ contains a total of 254 columns.This matrix is created to be used in future research on quantifying of meaning in scientific texts under the assumption that words have scientifically specific meanings in subject categories and the meaning can be estimated by information gains from word to categories. LScT (Leicester Scientific Thesaurus) is a scientific thesaurus of English. The thesaurus includes a list of 5,000 words from the LScDC. We consider ordering the words of LScDC by the sum of their RIGs in categories. That is, words are arranged in their informativeness in the scientific corpus LSC. Therefore, meaningfulness of words evaluated by words’ average informativeness in the categories. We have decided to include the most informative 5,000 words in the scientific thesaurus. Words as a Vector of Frequencies in WoS CategoriesEach word of the LScDC is represented as a vector of frequencies in WoS categories. Given the collection of the LSC texts, each entry of the vector consists of the number of texts containing the word in the corresponding category.It is noteworthy that texts in a corpus do not necessarily belong to a single category, as they are likely to correspond to multidisciplinary studies, specifically in a corpus of scientific texts. In other words, categories may not be exclusive. There are 252 WoS categories and a text can be assigned to at least 1 and at most 6 categories in the LSC. Using the binary calculation of frequencies, we introduce the presence of a word in a category. We create a vector of frequencies for each word, where dimensions are categories in the corpus.The collection of vectors, with all words and categories in the entire corpus, can be shown in a table, where each entry corresponds to a pair (word,category). This table is build for the LScDC with 252 WoS categories and presented in published archive with this file. The value of each entry in the table shows how many times a word of LScDC appears in a WoS category. The occurrence of a word in a category is determined by counting the number of the LSC texts containing the word in a category. Words as a Vector of Relative Information Gains Extracted for CategoriesIn this section, we introduce our approach to representation of a word as a vector of relative information gains for categories under the assumption that meaning of a word can be quantified by their information gained for categories.For each category, a function is defined on texts that takes the value 1, if the text belongs to the category, and 0 otherwise. For each word, a function is defined on texts that takes the value 1 if the word belongs to the text, and 0 otherwise. Consider LSC as a probabilistic sample space (the space of equally probable elementary outcomes). For the Boolean random variables, the joint probability distribution, the entropy and information gains are defined.The information gain about the category from the word is the amount of information on the belonging of a text from the LSC to the category from observing the word in the text [6]. We used the Relative Information Gain (RIG) providing a normalised measure of the Information Gain. This provides the ability of comparing information gains for different categories. The calculations of entropy, Information Gains and Relative Information Gains can be found in the README file in the archive published. Given a word, we created a vector where each component of the vector corresponds to a category. Therefore, each word is represented as a vector of relative information gains. It is obvious that the dimension of vector for each word is the number of categories. The set of vectors is used to form the Word-Category RIG Matrix, in which each column corresponds to a category, each row corresponds to a word and each component is the relative information gain from the word to the category. In Word-Category RIG Matrix, a row vector represents the corresponding word as a vector of RIGs in categories. We note that in the matrix, a column vector represents RIGs of all words in an individual category. If we choose an arbitrary category, words can be ordered by their RIGs from the most informative to the least informative for the category. As well as ordering words in each category, words can be ordered by two criteria: sum and maximum of RIGs in categories. The top n words in this list can be considered as the most informative words in the scientific texts. For a given word, the sum and maximum of RIGs are calculated from the Word-Category RIG Matrix.RIGs for each word of LScDC in 252 categories are calculated and vectors of words are formed. We then form the Word-Category RIG Matrix for the LSC. For each word, the sum (S) and maximum (M) of RIGs in categories are calculated and added at the end of the matrix (last two columns of the matrix). The Word-Category RIG Matrix for the LScDC with 252 categories, the sum of RIGs in categories and the maximum of RIGs over categories can be found in the database.Leicester Scientific Thesaurus (LScT)Leicester Scientific Thesaurus (LScT) is a list of 5,000 words form the LScDC [2]. Words of LScDC are sorted in descending order by the sum (S) of RIGs in categories and the top 5,000 words are selected to be included in the LScT. We consider these 5,000 words as the most meaningful words in the scientific corpus. In other words, meaningfulness of words evaluated by words’ average informativeness in the categories and the list of these words are considered as a ‘thesaurus’ for science. The LScT with value of sum can be found as CSV file with the published archive. Published archive contains following files:1) Word_Category_RIG_Matrix.csv: A 103,998 by 254 matrix where columns are 252 WoS categories, the sum (S) and the maximum (M) of RIGs in categories (last two columns of the matrix), and rows are words of LScDC. Each entry in the first 252 columns is RIG from the word to the category. Words are ordered as in the LScDC.2) Word_Category_Frequency_Matrix.csv: A 103,998 by 252 matrix where columns are 252 WoS categories and rows are words of LScDC. Each entry of the matrix is the number of texts containing the word in the corresponding category. Words are ordered as in the LScDC.3) LScT.csv: List of words of LScT with sum (S) values. 4) Text_No_in_Cat.csv: The number of texts in categories. 5) Categories_in_Documents.csv: List of WoS categories for each document of the LSC.6) README.txt: Description of Word-Category RIG Matrix, Word-Category Frequency Matrix and LScT and forming procedures.7) README.pdf (same as 6 in PDF format)References[1] Suzen, Neslihan (2019): LSC (Leicester Scientific Corpus). figshare. Dataset. https://doi.org/10.25392/leicester.data.9449639.v2[2] Suzen, Neslihan (2019): LScDC (Leicester Scientific Dictionary-Core). figshare. Dataset. https://doi.org/10.25392/leicester.data.9896579.v3[3] Web of Science. (15 July). Available: https://apps.webofknowledge.com/[4] WoS Subject Categories. Available: https://images.webofknowledge.com/WOKRS56B5/help/WOS/hp_subject_category_terms_tasca.html [5] Suzen, N., Mirkes, E. M., & Gorban, A. N. (2019). LScDC-new large scientific dictionary. arXiv preprint arXiv:1912.06858. [6] Shannon, C. E. (1948). A mathematical theory of communication. Bell system technical journal, 27(3), 379-423.

LScD (Leicester Scientific 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 LScD (Leicester Scientific Dictionary) is created from the updated LSC (Leicester Scientific Corpus) - Version 2*. All pre-processing steps applied to build the new version of the dictionary are the same as in Version 2** and can be found in description of Version 2 below. We did not repeat the explanation. After pre-processing steps, the total number of unique words in the new version of the dictionary is 972,060. The files provided with this description are also same as described as for LScD Version 2 below.* Suzen, Neslihan (2019): LSC (Leicester Scientific Corpus). figshare. Dataset. https://doi.org/10.25392/leicester.data.9449639.v2** Suzen, Neslihan (2019): LScD (Leicester Scientific Dictionary). figshare. Dataset. https://doi.org/10.25392/leicester.data.9746900.v2[Version 2] Getting StartedThis document provides the pre-processing steps for creating an ordered list of words from the LSC (Leicester Scientific Corpus) [1] and the description of LScD (Leicester Scientific Dictionary). This dictionary is created to be used in future work on the quantification of the meaning of research texts. R code for producing the dictionary from LSC and instructions for usage of the code are available in [2]. The code can be also used for list of texts from other sources, amendments to the code may be required.LSC is a collection of abstracts of articles and proceeding papers published in 2014 and indexed by the Web of Science (WoS) database [3]. Each document contains title, list of authors, list of categories, list of research areas, and times cited. The corpus contains only documents in English. The corpus was collected in July 2018 and contains the number of citations from publication date to July 2018. The total number of documents in LSC is 1,673,824.LScD is an ordered list of words from texts of abstracts in LSC.The dictionary stores 974,238 unique words, is sorted by the number of documents containing the word in descending order. All words in the LScD are in stemmed form of words. The LScD contains the following information:1.Unique words in abstracts2.Number of documents containing each word3.Number of appearance of a word in the entire corpusProcessing the LSCStep 1.Downloading the LSC Online: Use of the LSC is subject to acceptance of request of the link by email. To access the LSC for research purposes, please email to ns433@le.ac.uk. The data are extracted from Web of Science [3]. You may not copy or distribute these data in whole or in part without the written consent of Clarivate Analytics.Step 2.Importing the Corpus to R: The full R code for processing the corpus can be found in the GitHub [2].All following steps can be applied for arbitrary list of texts from any source with changes of parameter. The structure of the corpus such as file format and names (also the position) of fields should be taken into account to apply our code. The organisation of CSV files of LSC is described in README file for LSC [1].Step 3.Extracting Abstracts and Saving Metadata: Metadata that include all fields in a document excluding abstracts and the field of abstracts are separated. Metadata are then saved as MetaData.R. Fields of metadata are: List_of_Authors, Title, Categories, Research_Areas, Total_Times_Cited and Times_cited_in_Core_Collection.Step 4.Text Pre-processing Steps on the Collection of Abstracts: In this section, we presented our approaches to pre-process abstracts of the LSC.1.Removing punctuations and special characters: This is the process of substitution of all non-alphanumeric characters by space. We did not substitute the character “-” in this step, because we need to keep words like “z-score”, “non-payment” and “pre-processing” in order not to lose the actual meaning of such words. A processing of uniting prefixes with words are performed in later steps of pre-processing.2.Lowercasing the text data: Lowercasing is performed to avoid considering same words like “Corpus”, “corpus” and “CORPUS” differently. Entire collection of texts are converted to lowercase.3.Uniting prefixes of words: Words containing prefixes joined with character “-” are united as a word. The list of prefixes united for this research are listed in the file “list_of_prefixes.csv”. The most of prefixes are extracted from [4]. We also added commonly used prefixes: ‘e’, ‘extra’, ‘per’, ‘self’ and ‘ultra’.4.Substitution of words: Some of words joined with “-” in the abstracts of the LSC require an additional process of substitution to avoid losing the meaning of the word before removing the character “-”. Some examples of such words are “z-test”, “well-known” and “chi-square”. These words have been substituted to “ztest”, “wellknown” and “chisquare”. Identification of such words is done by sampling of abstracts form LSC. The full list of such words and decision taken for substitution are presented in the file “list_of_substitution.csv”.5.Removing the character “-”: All remaining character “-” are replaced by space.6.Removing numbers: All digits which are not included in a word are replaced by space. All words that contain digits and letters are kept because alphanumeric characters such as chemical formula might be important for our analysis. Some examples are “co2”, “h2o” and “21st”.7.Stemming: Stemming is the process of converting inflected words into their word stem. This step results in uniting several forms of words with similar meaning into one form and also saving memory space and time [5]. All words in the LScD are stemmed to their word stem.8.Stop words removal: Stop words are words that are extreme common but provide little value in a language. Some common stop words in English are ‘I’, ‘the’, ‘a’ etc. We used ‘tm’ package in R to remove stop words [6]. There are 174 English stop words listed in the package.Step 5.Writing the LScD into CSV Format: There are 1,673,824 plain processed texts for further analysis. All unique words in the corpus are extracted and written in the file “LScD.csv”.The Organisation of the LScDThe total number of words in the file “LScD.csv” is 974,238. Each field is described below:Word: It contains unique words from the corpus. All words are in lowercase and their stem forms. The field is sorted by the number of documents that contain words in descending order.Number of Documents Containing the Word: In this content, binary calculation is used: if a word exists in an abstract then there is a count of 1. If the word exits more than once in a document, the count is still 1. Total number of document containing the word is counted as the sum of 1s in the entire corpus.Number of Appearance in Corpus: It contains how many times a word occurs in the corpus when the corpus is considered as one large document.Instructions for R CodeLScD_Creation.R is an R script for processing the LSC to create an ordered list of words from the corpus [2]. Outputs of the code are saved as RData file and in CSV format. Outputs of the code are:Metadata File: It includes all fields in a document excluding abstracts. Fields are List_of_Authors, Title, Categories, Research_Areas, Total_Times_Cited and Times_cited_in_Core_Collection.File of Abstracts: It contains all abstracts after pre-processing steps defined in the step 4.DTM: It is the Document Term Matrix constructed from the LSC[6]. Each entry of the matrix is the number of times the word occurs in the corresponding document.LScD: An ordered list of words from LSC as defined in the previous section.The code can be used by:1.Download the folder ‘LSC’, ‘list_of_prefixes.csv’ and ‘list_of_substitution.csv’2.Open LScD_Creation.R script3.Change parameters in the script: replace with the full path of the directory with source files and the full path of the directory to write output files4.Run the full code.References[1]N. Suzen. (2019). LSC (Leicester Scientific Corpus) [Dataset]. Available: https://doi.org/10.25392/leicester.data.9449639.v1[2]N. Suzen. (2019). LScD-LEICESTER SCIENTIFIC DICTIONARY CREATION. Available: https://github.com/neslihansuzen/LScD-LEICESTER-SCIENTIFIC-DICTIONARY-CREATION[3]Web of Science. (15 July). Available: https://apps.webofknowledge.com/[4]A. Thomas, "Common Prefixes, Suffixes and Roots," Center for Development and Learning, 2013.[5]C. Ramasubramanian and R. Ramya, "Effective pre-processing activities in text mining using improved porter’s stemming algorithm," International Journal of Advanced Research in Computer and Communication Engineering, vol. 2, no. 12, pp. 4536-4538, 2013.[6]I. Feinerer, "Introduction to the tm Package Text Mining in R," Accessible en ligne: https://cran.r-project.org/web/packages/tm/vignettes/tm.pdf, 2013.

Excelfile that contains data on loanwords and native words in the Nordic languages during the period 1550 to 1900.

This data stems from a two-years research about loanwords and native words in the Nordic languages (Swedish, Danish, Icelandic) in the period c. 1550-c. 1900. The data gathering focused on a set of meanings from different semantic fields in the quest for loanwords and native words throughout time. The semantic fields from which data was sampled are:

Astrology Astronomy Buildings and Architectural Elements Chemical Elements Clothes and Accessories Food and Beverages Grammatical Terminology Law Learning Materials, Minerals, Textiles and Fabrics Mathematics and Geometry Medicine Musical Instruments Natural Products and Plants People and Honorific and Professional Titles Religious Terminology Statal Organization

The dataset was originally published in DiVA and moved to SND in 2024.

HindEnCorp parallel texts (sentence-aligned) come from the following sources: Tides, which contains 50K sentence pairs taken mainly from news articles. This dataset was originally col- lected for the DARPA-TIDES surprise-language con- test in 2002, later refined at IIIT Hyderabad and provided for the NLP Tools Contest at ICON 2008 (Venkatapathy, 2008).

Commentaries by Daniel Pipes contain 322 articles in English written by a journalist Daniel Pipes and translated into Hindi.

EMILLE. This corpus (Baker et al., 2002) consists of three components: monolingual, parallel and annotated corpora. There are fourteen monolingual sub- corpora, including both written and (for some lan- guages) spoken data for fourteen South Asian lan- guages. The EMILLE monolingual corpora contain in total 92,799,000 words (including 2,627,000 words of transcribed spoken data for Bengali, Gujarati, Hindi, Punjabi and Urdu). The parallel corpus consists of 200,000 words of text in English and its accompanying translations into Hindi and other languages.

Smaller datasets as collected by Bojar et al. (2010) include the corpus used at ACL 2005 (a subcorpus of EMILLE), a corpus of named entities from Wikipedia (crawled in 2009), and Agriculture domain parallel corpus.  For the current release, we are extending the parallel corpus using these sources: Intercorp (Čermák and Rosen,2012) is a large multilingual parallel corpus of 32 languages including Hindi. The central language used for alignment is Czech. Intercorp’s core texts amount to 202 million words. These core texts are most suitable for us because their sentence alignment is manually checked and therefore very reliable. They cover predominately short sto- ries and novels. There are seven Hindi texts in Inter- corp. Unfortunately, only for three of them the English translation is available; the other four are aligned only with Czech texts. The Hindi subcorpus of Intercorp contains 118,000 words in Hindi.

TED talks 3 held in various languages, primarily English, are equipped with transcripts and these are translated into 102 languages. There are 179 talks for which Hindi translation is available.

The Indic multi-parallel corpus (Birch et al., 2011; Post et al., 2012) is a corpus of texts from Wikipedia translated from the respective Indian language into English by non-expert translators hired over Mechanical Turk. The quality is thus somewhat mixed in many respects starting from typesetting and punctuation over capi- talization, spelling, word choice to sentence structure. A little bit of control could be in principle obtained from the fact that every input sentence was translated 4 times. We used the 2012 release of the corpus.

Launchpad.net is a software collaboration platform that hosts many open-source projects and facilitates also collaborative localization of the tools. We downloaded all revisions of all the hosted projects and extracted the localization (.po) files.

Other smaller datasets. This time, we added Wikipedia entities as crawled in 2013 (including any morphological variants of the named entitity that appears on the Hindi variant of the Wikipedia page) and words, word examples and quotes from the Shabdkosh online dictionary.

TripAdvisor reviews and comparable data sources play an important role in many tasks in Natural Language Processing (NLP), providing a data basis for the identification and classification of subjective judgments, such as hotel or restaurant reviews, into positive or negative polarities. This study explores three important factors influencing variation in crowdsourced polarity judgments, focusing on TripAdvisor reviews in Spanish. Three hypotheses are tested: the role of Part Of Speech (POS), the impact of sentiment words such as “tasty”, and the influence of neutral words like “ok” on judgment variation. The study’s methodology employs one-word titles, demonstrating their efficacy in studying polarity variation of words. Statistical tests on mean equality are performed on word groups of our interest. The results of this study reveal that adjectives in one-word titles tend to result in lower judgment variation compared to other word types or POS. Sentiment words contribute to lower judgment variation as well, emphasizing the significance of sentiment words in research on polarity judgments, and neutral words are associated with higher judgment variation as expected. However, these effects cannot be always reproduced in longer titles, which suggests that longer titles do not represent the best data source for testing the ambiguity of single words due to the influence on word polarity by other words like negation in longer titles. This empirical investigation contributes valuable insights into the factors influencing polarity variation of words, providing a foundation for NLP practitioners that aim to capture and predict polarity judgments in Spanish and for researchers that aim to understand factors influencing judgment variation.

This data set is comprised of 300 target words, paired with taxonomically related and/or thematically related words for a total of 659 pairs. Each pair has been normed through online surveys for both their taxonomic similarity and thematic relatedness.

Context

WLASL is the largest video dataset for Word-Level American Sign Language (ASL) recognition, which features 2,000 common different words in ASL. We hope WLASL will facilitate the research in sign language understanding and eventually benefit the communication between deaf and hearing communities.

Content

The WLASL_v0.3.json file contains the glossary and instances of the videos.

Inside the videos folder, there are about 12k videos each named corresponding video_id.

Acknowledgements

All the WLASL data is intended for academic and computational use only. No commercial usage is allowed.

Made by Dongxu Li and Hongdong Li. Please read the WLASL paper and visit the official website and repository.

Licensed under the Computational Use of Data Agreement (C-UDA). Please refer to the C-UDA-1.0 page for more information.

Inspiration

• How to classify word-level action recognition to text?
• What is the most accurate model to do word-level sign language recognition?

As there was no large publicly available cross-domain dataset for comparative argument mining, we create one composed of sentences, potentially annotated with BETTER / WORSE markers (the first object is better / worse than the second object) or NONE (the sentence does not contain a comparison of the target objects). The BETTER sentences stand for a pro-argument in favor of the first compared object and WORSE-sentences represent a con-argument and favor the second object. We aim for minimizing dataset domain-specific biases in order to capture the nature of comparison and not the nature of the particular domains, thus decided to control the specificity of domains by the selection of comparison targets. We hypothesized and could confirm in preliminary experiments that comparison targets usually have a common hypernym (i.e., are instances of the same class), which we utilized for selection of the compared objects pairs. The most specific domain we choose, is computer science with comparison targets like programming languages, database products and technology standards such as Bluetooth or Ethernet. Many computer science concepts can be compared objectively (e.g., on transmission speed or suitability for certain applications). The objects for this domain were manually extracted from List of-articles at Wikipedia. In the annotation process, annotators were asked to only label sentences from this domain if they had some basic knowledge in computer science. The second, broader domain is brands. It contains objects of different types (e.g., cars, electronics, and food). As brands are present in everyday life, anyone should be able to label the majority of sentences containing well-known brands such as Coca-Cola or Mercedes. Again, targets for this domain were manually extracted from `List of''-articles at Wikipedia.The third domain is not restricted to any topic: random. For each of 24~randomly selected seed words 10 similar words were collected based on the distributional similarity API of JoBimText (http://www.jobimtext.org). Seed words created using randomlists.com: book, car, carpenter, cellphone, Christmas, coffee, cork, Florida, hamster, hiking, Hoover, Metallica, NBC, Netflix, ninja, pencil, salad, soccer, Starbucks, sword, Tolkien, wine, wood, XBox, Yale.Especially for brands and computer science, the resulting object lists were large (4493 in brands and 1339 in computer science). In a manual inspection, low-frequency and ambiguous objects were removed from all object lists (e.g., RAID (a hardware concept) and Unity (a game engine) are also regularly used nouns). The remaining objects were combined to pairs. For each object type (seed Wikipedia list page or the seed word), all possible combinations were created. These pairs were then used to find sentences containing both objects. The aforementioned approaches to selecting compared objects pairs tend minimize inclusion of the domain specific data, but do not solve the problem fully though. We keep open a question of extending dataset with diverse object pairs including abstract concepts for future work. As for the sentence mining, we used the publicly available index of dependency-parsed sentences from the Common Crawl corpus containing over 14 billion English sentences filtered for duplicates. This index was queried for sentences containing both objects of each pair. For 90% of the pairs, we also added comparative cue words (better, easier, faster, nicer, wiser, cooler, decent, safer, superior, solid, terrific, worse, harder, slower, poorly, uglier, poorer, lousy, nastier, inferior, mediocre) to the query in order to bias the selection towards comparisons but at the same time admit comparisons that do not contain any of the anticipated cues. This was necessary as a random sampling would have resulted in only a very tiny fraction of comparisons. Note that even sentences containing a cue word do not necessarily express a comparison between the desired targets (dog vs. cat: He's the best pet that you can get, better than a dog or cat.). It is thus especially crucial to enable a classifier to learn not to rely on the existence of clue words only (very likely in a random sample of sentences with very few comparisons). For our corpus, we keep pairs with at least 100 retrieved sentences.From all sentences of those pairs, 2500 for each category were randomly sampled as candidates for a crowdsourced annotation that we conducted on figure-eight.com in several small batches. Each sentence was annotated by at least five trusted workers. We ranked annotations by confidence, which is the figure-eight internal measure of combining annotator trust and voting, and discarded annotations with a confidence below 50%. Of all annotated items, 71% received unanimous votes and for over 85% at least 4 out of 5 workers agreed -- rendering the collection procedure aimed at ease of annotation successful.The final dataset contains 7199 sentences with 271 distinct object pairs. The majority of sentences (over 72%) are non-comparative despite biasing the selection with cue words; in 70% of the comparative sentences, the favored target is named first.You can browse though the data here: https://docs.google.com/spreadsheets/d/1U8i6EU9GUKmHdPnfwXEuBxi0h3aiRCLPRC-3c9ROiOE/edit?usp=sharing Full description of the dataset is available in the workshop paper at ACL 2019 conference. Please cite this paper if you use the data: Franzek, Mirco, Alexander Panchenko, and Chris Biemann. ""Categorization of Comparative Sentences for Argument Mining."" arXiv preprint arXiv:1809.06152 (2018).@inproceedings{franzek2018categorization, title={Categorization of Comparative Sentences for Argument Mining}, author={Panchenko, Alexander and Bondarenko, and Franzek, Mirco and Hagen, Matthias and Biemann, Chris}, booktitle={Proceedings of the 6th Workshop on Argument Mining at ACL'2019}, year={2019}, address={Florence, Italy}}

Dataset

This dataset was created by Sang-eon Park

Released under Other (specified in description)

Contents

The Wake Word English Dataset is a comprehensive collection of audio samples designed to facilitate the training and evaluation of wake word detection systems. This dataset is tailored for the development of robust and accurate voice-activated systems, commonly used in virtual assistants, smart home devices, and various voice-controlled applications.

The LIST corpus extraction tool is a Java program for extracting lists from text corpora on the levels of characters, word parts, words, and word sets. It supports VERT and TEI P5 XML formats and outputs .CSV files that can be imported into Microsoft Excel or similar statistical processing software.

Version 1.3 adds support for the KOST 2.0 Slovene Learner Corpus (http://hdl.handle.net/11356/1887) in XML format. It also allows program execution using the command line (see 00README.txt for details), and uses a later version of Java (tested using JDK 21). In addition, Windows users no longer need to have Java installed on their computers to run the program.

Thesaurus of Modern Slovene is the largest automatically generated open-access collection of Slovene synonyms. It is sourced from the data in two principal language resources: The Oxford®-DZS Comprehensive English-Slovenian Dictionary and the Gigafida 1.0 corpus of written Slovene. The links identified between synonyms were additionally confirmed using the Dictionary of Standard Slovenian Language (SSKJ). The data extraction and structure for the Thesaurus were based on the frequency and manner in which words co-occur in translation strings of the Oxford-DZS Dictionary. This information is the basis for discriminating between ‘core’ and ‘near’ synonyms, with ‘core’ synonyms exhibiting a greater connection to the keyword. In the following step, an approach combining balanced co-occurrence graphs and the Personal PageRank algorithm automatically divides the synonyms into subgroups and ranks them according to the degree of semantic relatedness to the keyword, as well as their frequency in language use. For the creation methodology, see Krek et al. (2017) in the provided references.

The database includes dictionary entries: single- and multiword headwords, their part-of-speech and other linguistic features, as well as automatically extracted synonyms, their type (core or near) and relevancy rank. In version 2.0, 4,544 manually revised antonyms were added to the database. Additionally, for a part of the database, synonyms were distributed under the corresponding word senses. Pertaining to how much lexicographic revision was involved in their preparation, database entries can have one of the following three statuses: (a) ssss-automatic (96,064 entries): no manual revision was conducted; (b) ssss-manual (3,421 entries): word senses and semantic indicators were prepared by lexicographers, and synonyms were manually distributed under each corresponding sense; (c) ssss-hybrid (1,352 entries): manually revised senses are combined with data compiled automatically. For novelties of v2.0, see Arhar Holdt et al. (2023) in the provided references.

Introduction

Welcome to the US English Wake Word & Command Dataset, meticulously designed to advance the development and accuracy of voice-activated systems. This dataset features an extensive collection of wake words and commands, essential for triggering and interacting with voice assistants and other voice-activated devices. Our dataset ensures these systems respond promptly and accurately to user inputs, enhancing their reliability and user experience.

Speech Data

This training dataset comprises over 20,000 audio recordings of wake words and command phrases designed to build robust and accurate voice assistant speech technology. Each participant recorded 400 recordings in diverse environments and at varying speeds. This dataset contains audio recordings of wake words, as well as wake words followed by commands.

Dataset Diversity

This dataset includes recordings of various types of wake words and commands, in different environments and at different speeds, making it highly diverse.

This extensive coverage ensures the dataset includes realistic scenarios, which is essential for developing effective voice assistant speech recognition models.

Metadata

The dataset provides comprehensive metadata for each audio recording and participant:

Introduction

There are several works based on Natural Language Processing on newspaper reports. Mining opinions from headlines [ 1 ] using Standford NLP and SVM by Rameshbhaiet. Al.compared several algorithms on a small and large dataset. Rubinet. al., in their paper [ 2 ], created a mechanism to differentiate fake news from real ones by building a set of characteristics of news according to their types. The purpose was to contribute to the low resource data available for training machine learning algorithms. Doumitet. al.in [ 3 ] have implemented LDA, a topic modeling approach to study bias present in online news media.

However, there are not many NLP research invested in studying COVID-19. Most applications include classification of chest X-rays and CT-scans to detect presence of pneumonia in lungs [ 4 ], a consequence of the virus. Other research areas include studying the genome sequence of the virus[ 5 ][ 6 ][ 7 ] and replicating its structure to fight and find a vaccine. This research is crucial in battling the pandemic. The few NLP based research publications are sentiment classification of online tweets by Samuel et el [ 8 ] to understand fear persisting in people due to the virus. Similar work has been done using the LSTM network to classify sentiments from online discussion forums by Jelodaret. al.[ 9 ]. NKK dataset is the first study on a comparatively larger dataset of a newspaper report on COVID-19, which contributed to the virus’s awareness to the best of our knowledge.

2 Data-set Introduction

2.1 Data Collection

We accumulated 1000 online newspaper report from United States of America (USA) on COVID-19. The newspaper includes The Washington Post (USA) and StarTribune (USA). We have named it as “Covid-News-USA-NNK”. We also accumulated 50 online newspaper report from Bangladesh on the issue and named it “Covid-News-BD-NNK”. The newspaper includes The Daily Star (BD) and Prothom Alo (BD). All these newspapers are from the top provider and top read in the respective countries. The collection was done manually by 10 human data-collectors of age group 23- with university degrees. This approach was suitable compared to automation to ensure the news were highly relevant to the subject. The newspaper online sites had dynamic content with advertisements in no particular order. Therefore there were high chances of online scrappers to collect inaccurate news reports. One of the challenges while collecting the data is the requirement of subscription. Each newspaper required $1 per subscriptions. Some criteria in collecting the news reports provided as guideline to the human data-collectors were as follows:

The headline must have one or more words directly or indirectly related to COVID-19.

The content of each news must have 5 or more keywords directly or indirectly related to COVID-19.

The genre of the news can be anything as long as it is relevant to the topic. Political, social, economical genres are to be more prioritized.

Avoid taking duplicate reports.

Maintain a time frame for the above mentioned newspapers.

To collect these data we used a google form for USA and BD. We have two human editor to go through each entry to check any spam or troll entry.

2.2 Data Pre-processing and Statistics

Some pre-processing steps performed on the newspaper report dataset are as follows:

Remove hyperlinks.

Remove non-English alphanumeric characters.

Remove stop words.

Lemmatize text.

While more pre-processing could have been applied, we tried to keep the data as much unchanged as possible since changing sentence structures could result us in valuable information loss. While this was done with help of a script, we also assigned same human collectors to cross check for any presence of the above mentioned criteria.

The primary data statistics of the two dataset are shown in Table 1 and 2.

Table 1: Covid-News-USA-NNK data statistics

No of words per headline

7 to 20

No of words per body content

150 to 2100

Table 2: Covid-News-BD-NNK data statistics No of words per headline

10 to 20

No of words per body content

100 to 1500

2.3 Dataset Repository

We used GitHub as our primary data repository in account name NKK^1. Here, we created two repositories USA-NKK^2 and BD-NNK^3. The dataset is available in both CSV and JSON format. We are regularly updating the CSV files and regenerating JSON using a py script. We provided a python script file for essential operation. We welcome all outside collaboration to enrich the dataset.

3 Literature Review

Natural Language Processing (NLP) deals with text (also known as categorical) data in computer science, utilizing numerous diverse methods like one-hot encoding, word embedding, etc., that transform text to machine language, which can be fed to multiple machine learning and deep learning algorithms.

Some well-known applications of NLP includes fraud detection on online media sites[ 10 ], using authorship attribution in fallback authentication systems[ 11 ], intelligent conversational agents or chatbots[ 12 ] and machine translations used by Google Translate[ 13 ]. While these are all downstream tasks, several exciting developments have been made in the algorithm solely for Natural Language Processing tasks. The two most trending ones are BERT[ 14 ], which uses bidirectional encoder-decoder architecture to create the transformer model, that can do near-perfect classification tasks and next-word predictions for next generations, and GPT-3 models released by OpenAI[ 15 ] that can generate texts almost human-like. However, these are all pre-trained models since they carry huge computation cost. Information Extraction is a generalized concept of retrieving information from a dataset. Information extraction from an image could be retrieving vital feature spaces or targeted portions of an image; information extraction from speech could be retrieving information about names, places, etc[ 16 ]. Information extraction in texts could be identifying named entities and locations or essential data. Topic modeling is a sub-task of NLP and also a process of information extraction. It clusters words and phrases of the same context together into groups. Topic modeling is an unsupervised learning method that gives us a brief idea about a set of text. One commonly used topic modeling is Latent Dirichlet Allocation or LDA[17].

Keyword extraction is a process of information extraction and sub-task of NLP to extract essential words and phrases from a text. TextRank [ 18 ] is an efficient keyword extraction technique that uses graphs to calculate the weight of each word and pick the words with more weight to it.

Word clouds are a great visualization technique to understand the overall ’talk of the topic’. The clustered words give us a quick understanding of the content.

4 Our experiments and Result analysis

We used the wordcloud library^4 to create the word clouds. Figure 1 and 3 presents the word cloud of Covid-News-USA- NNK dataset by month from February to May. From the figures 1,2,3, we can point few information:

In February, both the news paper have talked about China and source of the outbreak.

StarTribune emphasized on Minnesota as the most concerned state. In April, it seemed to have been concerned more.

Both the newspaper talked about the virus impacting the economy, i.e, bank, elections, administrations, markets.

Washington Post discussed global issues more than StarTribune.

StarTribune in February mentioned the first precautionary measurement: wearing masks, and the uncontrollable spread of the virus throughout the nation.

While both the newspaper mentioned the outbreak in China in February, the weight of the spread in the United States are more highlighted through out March till May, displaying the critical impact caused by the virus.

We used a script to extract all numbers related to certain keywords like ’Deaths’, ’Infected’, ’Died’ , ’Infections’, ’Quarantined’, Lock-down’, ’Diagnosed’ etc from the news reports and created a number of cases for both the newspaper. Figure 4 shows the statistics of this series. From this extraction technique, we can observe that April was the peak month for the covid cases as it gradually rose from February. Both the newspaper clearly shows us that the rise in covid cases from February to March was slower than the rise from March to April. This is an important indicator of possible recklessness in preparations to battle the virus. However, the steep fall from April to May also shows the positive response against the attack. We used Vader Sentiment Analysis to extract sentiment of the headlines and the body. On average, the sentiments were from -0.5 to -0.9. Vader Sentiment scale ranges from -1(highly negative to 1(highly positive). There were some cases

where the sentiment scores of the headline and body contradicted each other,i.e., the sentiment of the headline was negative but the sentiment of the body was slightly positive. Overall, sentiment analysis can assist us sort the most concerning (most negative) news from the positive ones, from which we can learn more about the indicators related to COVID-19 and the serious impact caused by it. Moreover, sentiment analysis can also provide us information about how a state or country is reacting to the pandemic. We used PageRank algorithm to extract keywords from headlines as well as the body content. PageRank efficiently highlights important relevant keywords in the text. Some frequently occurring important keywords extracted from both the datasets are: ’China’, Government’, ’Masks’, ’Economy’, ’Crisis’, ’Theft’ , ’Stock market’ , ’Jobs’ , ’Election’, ’Missteps’, ’Health’, ’Response’. Keywords extraction acts as a filter allowing quick searches for indicators in case of locating situations of the economy,

The COVID-19 infodemic, characterized by the rapid spread of misinformation and unverified claims related to the pandemic, presents a significant challenge. This paper presents a comparative analysis of the COVID-19 infodemic in the English and Chinese languages, utilizing textual data extracted from social media platforms. To ensure a balanced representation, two infodemic datasets were created by augmenting previously collected social media textual data. Through word frequency analysis, the 30 most frequently occurring infodemic words are identified, shedding light on prevalent discussions surrounding the infodemic. Moreover, topic clustering analysis uncovers thematic structures and provides a deeper understanding of primary topics within each language context. Additionally, sentiment analysis enables comprehension of the emotional tone associated with COVID-19 information on social media platforms in English and Chinese. This research contributes to a better understanding of the COVID-19 infodemic phenomenon and can guide the development of strategies to combat misinformation during public health crises across different languages.

Open access platforms and retail websites have one thing in common: they are trying to present the most relevant offerings possible to their patrons. Retail websites – such as Amazon.com – de-ploy recommender systems based on data collected about their customers. These systems im-prove with the amount of data available: the more is known about the customers, the better it can predict what other merchandise will appeal.Recommender systems are successful, but using open access platforms to track people is not ac-ceptable. Therefore, a different solution is needed. Compared to retail websites, open access plat-forms have an unique advantage: they are able to use the complete contents of the publications they host. So, the question arises if it is possible to create a recommender system based on the contents of freely available documents, instead of personal data.The solution described in this paper is based on standard open source software. It is built using a combination of DSpace 6 and the R programming language. The open access platform – based on DSpace 6 – is the OAPEN Library; the data set used consists of nearly 11,000 open access books and chapters. The OAPEN Library enables data extraction through an API (application programming in-terface). A text mining algorithm written in the R programming language uses the full text of the publications and filters out the most common combinations of three words (trigrams). The next step is finding the publications that have one of more trigrams in common. The more trigrams two books or chapters share, the more closer they are 'connected'. This allows us not just to find relat-ed titles for each publication, but also to quantify how closely they are connected. Date: 2021-02-24 Date Submitted: 2021-02-24

Usage-based linguistics postulates that multi-word expressions constitute a substantial part of language structure and use, and are formed through repeated chunking and stored as exemplar wholes. They are also re-used to produce new sequences by means of schematization. While there is extensive research on multi-word expressions in many spoken languages, little is known about the status of multi-word expressions in the mainstream U.S. variety of American Sign Language (ASL). This paper investigates recurring multi-word expressions, or sequences of multiple signs, that involve a high-frequency sign of visual perception glossed as LOOK and the family of ‘look’ signs. The LOOK sign exhibits two broad functions: LOOK/‘vision’ references literal or metaphorical vision and LOOK/‘reaction’ signals a person’s reaction to a visual stimulus. Data analysis reveals that there are recurring sequences in distinct syntactic environments associated with the two functions of LOOK, suggesting that LOOK i...

The CHI-KNOW-PO (Chinese Knowledge and Poetry) database is structured around Chinese words and their use in the Medieval Period. As such, it contains a table of words, indications to which group they belong to (e.g. plants, emotions), what their synonyms are and the first known source to use the word. It contains a large number of other data related to this core: information about the source texts, their authors and commentators etc. In addition to words, semantic categories and synonyms, the database holds information on persons of Chinese history, literary works and connections between these objects. This part is linked to words (as we record first identified occurrence of words in texts by authors). It is nonetheless the result of a specific development. For more information, see: https://gitlab.huma-num.fr/chi-know-po/all-about-plants/-/wikis/home/database

In this repository, we release a series of vector space models of Ancient Greek, trained following different architectures and with different hyperparameter values.

Below is a breakdown of all the models released, with an indication of the training method and hyperparameters. The models are split into ‘Diachronica’ and ‘ALP’ models, according to the published paper they are associated with.

[Diachronica:] Stopponi, Silvia, Nilo Pedrazzini, Saskia Peels-Matthey, Barbara McGillivray & Malvina Nissim. Forthcoming. Natural Language Processing for Ancient Greek: Design, Advantages, and Challenges of Language Models, Diachronica.

[ALP:] Stopponi, Silvia, Nilo Pedrazzini, Saskia Peels-Matthey, Barbara McGillivray & Malvina Nissim. 2023. Evaluation of Distributional Semantic Models of Ancient Greek: Preliminary Results and a Road Map for Future Work. Proceedings of the Ancient Language Processing Workshop associated with the 14th International Conference on Recent Advances in Natural Language Processing (RANLP 2023). 49-58. Association for Computational Linguistics (ACL). https://doi.org/10.26615/978-954-452-087-8.2023_006

Diachronica models

Training data

Diorisis corpus (Vatri & McGillivray 2018). Separate models were trained for:

Classical subcorpus

Hellenistic subcorpus

Whole corpus

Models are named according to the (sub)corpus they are trained on (i.e. hel_ or hellenestic is appended to the name of the models trained on the Hellenestic subcorpus, clas_ or classical for the Classical subcorpus, full_ for the whole corpus).

Models

Count-based

Software used: LSCDetection (Kaiser et al. 2021; https://github.com/Garrafao/LSCDetection)

a. With Positive Pointwise Mutual Information applied (folder PPMI spaces). For each model, a version trained on each subcorpus after removing stopwords is also included (_stopfilt is appended to the model names). Hyperparameter values: window=5, k=1, alpha=0.75.

b. With both Positive Pointwise Mutual Information and dimensionality reduction with Singular Value Decomposition applied (folder PPMI+SVD spaces). For each model, a version trained on each subcorpus after removing stopwords is also included (_stopfilt is appended to the model names). Hyperparameter values: window=5, dimensions=300, gamma=0.0.

Word2Vec

Software used: CADE (Bianchi et al. 2020; https://github.com/vinid/cade).

a. Continuous-bag-of-words (CBOW). Hyperparameter values: size=30, siter=5, diter=5, workers=4, sg=0, ns=20.

b. Skipgram with Negative Sampling (SGNS). Hyperparameter values: size=30, siter=5, diter=5, workers=4, sg=1, ns=20.

Syntactic word embeddings

Syntactic word embeddings were also trained on the Ancient Greek subcorpus of the PROIEL treebank (Haug & Jøhndal 2008), the Gorman treebank (Gorman 2020), the PapyGreek treebank (Vierros & Henriksson 2021), the Pedalion treebank (Keersmaekers et al. 2019), and the Ancient Greek Dependency Treebank (Bamman & Crane 2011) largely following the SuperGraph method described in Al-Ghezi & Kurimo (2020) and the Node2Vec architecture (Grover & Leskovec 2016) (see https://github.com/npedrazzini/ancientgreek-syntactic-embeddings for more details). Hyperparameter values: window=1, min_count=1.

ALP models

Training data

Archaic, Classical, and Hellenistic portions of the Diorisis corpus (Vatri & McGillivray 2018) merged, stopwords removed according to the list made by Alessandro Vatri, available at https://figshare.com/articles/dataset/Ancient_Greek_stop_words/9724613.

Models

Count-based

Software used: LSCDetection (Kaiser et al. 2021; https://github.com/Garrafao/LSCDetection)

a. With Positive Pointwise Mutual Information applied (folder ppmi_alp). Hyperparameter values: window=5, k=1, alpha=0.75. Stopwords were removed from the training set.

b. With both Positive Pointwise Mutual Information and dimensionality reduction with Singular Value Decomposition applied (folder ppmi_svd_alp). Hyperparameter values: window=5, dimensions=300, gamma=0.0. Stopwords were removed from the training set.

Word2Vec

Software used: Gensim library (Řehůřek and Sojka, 2010)

a. Continuous-bag-of-words (CBOW). Hyperparameter values: size=30, window=5, min_count=5, negative=20, sg=0. Stopwords were removed from the training set.

b. Skipgram with Negative Sampling (SGNS). Hyperparameter values: size=30, window=5, min_count=5, negative=20, sg=1. Stopwords were removed from the training set.

References

Al-Ghezi, Ragheb & Mikko Kurimo. 2020. Graph-based syntactic word embeddings. In Ustalov, Dmitry, Swapna Somasundaran, Alexander Panchenko, Fragkiskos D. Malliaros, Ioana Hulpuș, Peter Jansen & Abhik Jana (eds.), Proceedings of the Graph-based Methods for Natural Language Processing (TextGraphs), 72-78.

Bamman, D. & Gregory Crane. 2011. The Ancient Greek and Latin dependency treebanks. In Sporleder, Caroline, Antal van den Bosch & Kalliopi Zervanou (eds.), Language Technology for Cultural Heritage. Selected Papers from the LaTeCH [Language Technology for Cultural Heritage] Workshop Series. Theory and Applications of Natural Language Processing, 79-98. Berlin, Heidelberg: Springer.

Gorman, Vanessa B. 2020. Dependency treebanks of Ancient Greek prose. Journal of Open Humanities Data 6(1).

Grover, Aditya & Jure Leskovec. 2016. Node2vec: scalable feature learning for networks. In Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD ‘16), 855-864.

Haug, Dag T. T. & Marius L. Jøhndal. 2008. Creating a parallel treebank of the Old Indo-European Bible translations. In Proceedings of the Second Workshop on Language Technology for Cultural Heritage Data (LaTeCH), 27–34.

Keersmaekers, Alek, Wouter Mercelis, Colin Swaelens & Toon Van Hal. 2019. Creating, enriching and valorizing treebanks of Ancient Greek. In Candito, Marie, Kilian Evang, Stephan Oepen & Djamé Seddah (eds.), Proceedings of the 18th International Workshop on Treebanks and Linguistic Theories (TLT, SyntaxFest 2019), 109-117.

Kaiser, Jens, Sinan Kurtyigit, Serge Kotchourko & Dominik Schlechtweg. 2021. Effects of Pre- and Post-Processing on type-based Embeddings in Lexical Semantic Change Detection. In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics.

Schlechtweg, Dominik, Anna Hätty, Marco del Tredici & Sabine Schulte im Walde. 2019. A Wind of Change: Detecting and Evaluating Lexical Semantic Change across Times and Domains. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, 732-746, Florence, Italy. ACL.

Vatri, Alessandro & Barbara McGillivray. 2018. The Diorisis Ancient Greek Corpus: Linguistics and Literature. Research Data Journal for the Humanities and Social Sciences 3, 1, 55-65, Available From: Brill https://doi.org/10.1163/24523666-01000013

Vierros, Marja & Erik Henriksson. 2021. PapyGreek treebanks: a dataset of linguistically annotated Greek documentary papyri. Journal of Open Humanities Data 7.