OSCAR or Open Super-large Crawled Aggregated coRpus is a huge multilingual corpus obtained by language classification and filtering of the Common Crawl corpus using the Ungoliant architecture.

This dataset represents a 2022 snapshot of the Supreme Court of Israel public verdicts and decisions supported by rich metadata. The 5.31GB dataset represents 751,194 documents. Overall, the dataset contains 2.68 Gb of text.

The texts in all of Hebrew Wikipedia was also extracted to pre-train OnlpLab's AlephBERT, using Attardi's Wikiextractor.

The National Library offers a collection of Jewish newspapers published in various countries, languages, and time periods, including digital versions and full-text search. The texts are published under a custom Terms of Use document that prohibits commercial use, and additionally requires checking the copyright status and receiving permission from the copyright-holder of the work for any use requiring such permission according to the Copyright Law.

A Hebrew Deduplicated and Cleaned Common Crawl Corpus. A thoroughly cleaned and approximately deduplicated dataset for unsupervised learning.

A corpus of 50K sentences from Hebrew Wikipedia chosen to ensure phoneme coverage for the purpose of a sentence recording project.

Structured Jewish texts and metadata with free public licenses, exported from Sefaria's database. A Living Library of Jewish Texts. 3,000 years of Jewish texts in Hebrew and English translation.

~15,000 israeli songs scrapped from Shironet website and contains 167 different singers. Contains only Hebrew characters.

1001 Israeli pop songs manual analyses 1967-2017.

Wikipedia, the free encyclopedia, publishes dumps of its content as XML files on a monthly basis.

Wikimedia dump service.

A repository containing dumps of thousands of public domain works in Hebrew, from Project Ben-Yehuda, in plaintext UTF-8 files, with and without diacritics (nikkud), and in HTML files.

This corpus is an attempt to recreate the dataset used for training XLM-R. This corpus comprises monolingual data for 100+ languages, including Hebrew. This was constructed using the urls and paragraph indices provided by the CC-Net repository by processing January-December 2018 Commoncrawl snapshots.

The HC Corpora was a great resource that contains natural language text from various newspapers, social media posts and blog pages in multiple languages. This is a cleaned version of the raw data from the newspaper subset of the HC corpus.

Hebrew textual corpus of construction, planning, and architecture, prepared by the BDAR Lab, Faculty of Architecture and Town Planning, Technion - IIT. The corpus consists of Hebrew documents from a wide variety of contemporary and historical sources, including legislative decrees, regulatory guidelines, research reports, academic studies, and professional journals. In the development of the corpus, it has been used digitally born as well as scanned printed publications, which go through a process of optical character recognition (OCR), cleaning, and parsing. This work was supported by the Israel Innovation Authority.

Transcripts and more in 12 languages including Hebrew

A corpus of transcriptions of Knesset (Israeli parliament) meetings between January 2004 and November 2005. Includes tokenized and morphologically tagged versions of most of the documents in the corpus.

The Hebrew Treebank Version 2.0 contains 6500 hand-annotated sentences of news items from the MILA HaAretz Corpus, with full word segmentation and morpho-syntactic analysis. Morphological features that are not directly relevant for syntactic structures, like roots, templates and patterns, are not analyzed. This resource can be used freely for research purposes only.

The Hebrew Universal Dependencies Treebank

This is the Modern Hebrew Dependency Treebank which was created and used in Yoav Goldberg's PhD thesis.

An annotated Hebrew database published as part of the Hebrew Language Corpus Project of Israel National Digital Agency and The Academy of the Hebrew Language.

Publicly available subset of the IAHLT UD Hebrew Treebank's Wikipedia section. The UD Hebrew-IAHLTWiki treebank consists of 5,000 contemporary Hebrew sentences representing a variety of texts originating from Wikipedia entries, compiled by the Israeli Association of Human Language Technology. It includes various text domains, such as: biography, law, finance, health, places, events and miscellaneous.

A Universal Dependencies treebank with named entities for contemporary Hebrew covering Knesset protocols.

IAHLT version of the UD Hebrew Treebank. This is a revised fork of the Universal Dependencies version of the Hebrew Treebank, with some important changes and a consistency overhaul involving substantial manual corrections. The dataset was prepared as part of the Hebrew & Arabic Corpus Linguistics Infrastructure project at the Israeli Association of Human Language Technologies (IAHLT).

The MILA center has 20 different corpora available for free for non-commercial use. All are available in plain text format, and most have tokenized, morphologically-analyzed, and morphologically-disambiguated versions available too. (temporarily down)

Named Entity (NER) annotations of the Hebrew Treebank (Haaretz newspaper) corpus, including: morpheme and token level NER labels, nested mentions, and more. The following entity types are tagged: Person, Organization, Geo-Political Entity, Location, Facility, Work-of-Art, Event, Product, Language.

A dataset of posts from the www.camoni.co.il, tagged with medical entities from the UMLS, and a code that recognize medical entities in the Hebrew text.

Newspaper articles in different fields: news, economy, fashion and gossip. The following entity types are tagged: entity names (person, location, organization), temporal expression (date, time) and number expression (percent, money).

a question answering dataset in Modern Hebrew, consisting of 30,147 questions. The dataset follows the format and crowdsourcing methodology of SQuAD (Stanford Question Answering Dataset) and the original ParaShoot. A team of crowdworkers formulated and answered reading comprehension questions based on random paragraphs in Hebrew. The answer to each question is a segment of text (span) included in the relevant paragraph. The paragraphs are sourced from two different platforms: (1) Hebrew Wikipedia, and (2) Geektime, an online Israeli news channel specializing in technology.

A Hebrew question and answering dataset in the style of SQuAD, created by Omri Keren and Omer Levy. ParaShoot is based on articles scraped from Wikipedia. The dataset contains 3K crowdsource-annotated pairs of questions and answers, in a setting suitable for few-shot learning.

Translated (by google translation API) SQUAD dataset from English to Hebrew. The translation process included fixation and removal of bad translations.

A sentiment analysis benchmark (positive, negative and neutral sentiment) for Hebrew, based on 12K social media comments, containing two instances of input items: token-based and morpheme-based. A cleaned version of the Hebrew Sentiment dataset - a test-train data leakage was cleaned.

A novel dataset for Hebrew Geo-Location, the first ever Hebrew NLU benchmark involving both grounding and geospatial reasoning, created with crowdsourced 5,649 geospatially-oriented Hebrew place descriptions of various place types from three cities in Israel.

Based on MultiNLI, a large crowd-sourced corpus of sentences from varied genres and writing styles in the English language. To adapt this resource for Hebrew, the corpus was translated from English using machine translation (Google Gemini). This dataset was created by Webiks for MAFAT, as part of the National Natural Language Processing Plan of Israel.

collected for HeBERT model and includes comments posted on news articles collected from 3 major Israeli news sites, between January 2020 to August 2020. The total size of the data is ~150 MB, including over 7 millions words and 350K sentences. ~2000 sentences were annotated by crowd members (3-10 annotators per sentence) for overall sentiment (polarity) and eight emotions: anger, disgust, expectation , fear, happiness, sadness, surprise and trust.

The sentiment analysis dataset contains 75,152 tagged sentences from 3 categories: economy, news (mostly politics) and sport. All the sentences were annotated by crowd members (2-5 annotators) to sentiment: positive, negative or neutral. This dataset was created by SUMIT-AI company, thanks to joint funding of the NNLP-IL.

This data was collected as a part of Nitzan Barzilay's project and contains about 2,700 quotes from Knesset meetings, manually classified into eight topics: education, Covid-19, welfare, economic, women and LGBT, health, security, internal security.

An archive of the writings of Zvi Yanai.

The corpus project, created by Dr Michal Marmorstein, Nadav Matalon, Amir Efrati, Itamar Folman and Yuval Geva, and hosted by the Hebrew University of Jerusalem (HUJI), aims at documenting naturally occurring speech and interaction in Modern Hebrew. Data come from telephone conversations recorded during the years 2020–2021. Data annotation followed standard methods of Interactional Linguistics (Couper-Kuhlen and Selting 2018). Audio files and transcripts were made freely accessible online.

Transcribed audio recordings (30 hours) of an Israeli economics podcast (חיות כיס).

A comprehensive Hebrew speech dataset designed for AI research and development. It contains approximately 3,300 hours of Hebrew speech, collected from a diverse range of online platforms including podcasts and other audio content.

A corpus of Hebrew dialogues within the Map Task framework (allowed for non-commercial research and teaching purposes only).

The Child Language Data Exchange System (CHILDES). Established to serve as a central repository for data of first language acquisition. Has contents (transcripts, audio, and video) in 26 languages from 230 different corpora

Natalie Shapira's large collection of Hebrew psychological lexicons and word lists. Useful for various psychology applications such as detecting emotional state, well being, relationship quality in conversation, identifying topics (e.g., family, work) and many more.

Sentiment analysis, the task of automatically detecting whether a piece of text is positive or negative, generally relies on a hand-curated list of words with positive sentiment (good, great, awesome) and negative sentiment (bad, gross, awful). This dataset contains both positive and negative sentiment lexicons for 81 languages.

Easy-to-use word-to-word translations for 3,564 language pairs. Hebrew is one of the 62 supported languages, and thus word-to-word translation to/from Hebrew is supported for 61 languages.

A corpus of digital vocalized Hebrew texts created by Eran Tomer as part of his Master thesis. The corpus is found in the ``resources`` folder.

This lexicon include two files: (1) The Verb Complements Lexicon - an Excel XLSX file. (2) A ZIP file of additional statistics - The results from both a morphologically analyzed corpus and a syntacticaly analyzed corpus, seperated according to the statiscal test. Also includes a summary of the results.

An Excel XLSX file containing 23,327 Hebrew tokens in descending order of frequency.

a list of the 500 most common words (stop words) computed from discussions from the Tapuz People website, on a variety of subjects.

Stop words in 28 languages, including Hebrew.

Created by Eran Tomer (erantom@gmail.com).

Lists of street, company, given and last names. Created by Guy Laybovitz.

1000 most frequent words in Hebrew tweets during (roughly) 2018.

Place Names in the Hebrew Script. An open, attestation based, historical database. Kima currently holds 27,239 Places, with 94,650 alternate variants of their names and 236,744 attestations of these variants.

Hebrew most common words by Twitter based on tweets from March 2018 to March 2019.

Useful word lists extracted from Hspell 1.4 by Eyal Gruss.

List of stop words in Hebrew produced by using Universal Dependencies of the The Israeli Association of Human Language Technologies (IAHLT).

Trained on Wikipedia using fastText. Comes in both the binary and text default formats of fastText: binary+text - https://dl.fbaipublicfiles.com/fasttext/vectors-wiki/wiki.he.zip, text - https://dl.fbaipublicfiles.com/fasttext/vectors-wiki/wiki.he.vec. In the text format, each line contains a word followed by its embedding; Each value is space separated; Words are ordered by their frequency in a descending order.

Trained on data from Twitter. Developed by Ron Shemesh in Bar-Ilan University's NLP lab under the instruction of Dr. Yoav Goldberg. Contains vectors for over 1.4M words (as of January 2018). Comes in a zip with two files: a text file with a word list and a NumPy array file (npy file).

Trained on the Hebrew CoNLL17 corpus using Word2Vec continuous skipgram, with a vecotor dimension of 100 and a window size of 10. The vocabulary includes 672,384 words.

Trained on the Hebrew CoNLL17 corpus using ELMO. **NOTE:** The link at the repository might not work. To download a concerete version of the Hebrew embeddings: http://vectors.nlpl.eu/repository/20/154.zip

Read more on https://www.oreilly.com/content/capturing-semantic-meanings-using-deep-learning/

supplies the frequency of occurrence of any Hebrew letter cluster (mean occurrence per million). The corpus was assembled throughout the year 2001, and consists of text downloaded from 914 editions of the three major daily online Hebrew newspapers (Haaretz, Maariv, and Yediot Acharonot). After removing abbreviations, single characters, forms with counts that are less than 3 (mostly typos), and splitting hyphenated forms (vast majority were two words), the corpus totals 554,270 types and 619,835,788 tokens.

About 255,000 inflect forms of 8,200 verbs that are used in contemporary Hebrew; About 280,000 declined forms of 15,000 nouns; about 22,000 lexical entries, including grammatical features and definitions. Made by Academy of the Hebrew Language (האקדמיה ללשון העברית).

LASER is a library to calculate and use multilingual sentence embeddings.

BERT, or Bidirectional Encoder Representations from Transformers, is a new method of pre-training language representations which obtains state-of-the-art results on a wide array of Natural Language Processing (NLP) tasks.

Iddo Yadlin and Itamar Shefi's word2vec model for Hebrew, trained on a corpus which is the Hebrew wikipedia dump only tokenized with hebpipe.

BERT Embeddings for Rabbinic-Encoded Language - DICTA's pre-trained language model (PLM) for Rabbinic Hebrew.

A very simple python tokenizer for Hebrew text.

Eyal Gruss's Hebrew tokenizer. A field-tested Hebrew tokenizer for dirty texts (ben-yehuda project, bible, cc100, mc4, opensubs, oscar, twitter) focused on multi-word expression extraction.

a BERT-style masked language model over character sequences, published by Omri Keren, Tal Avinari, Prof. Reut Tsarfaty and Dr. Omer Levy. Paper: https://arxiv.org/pdf/2204.04748.pdf

A highly accurate morphological segmenter to break up complex word forms.

A generation mechanism, created as part of Eran Tomer's (erantom@gmail.com) Master thesis, which produces vocalized and morphologically tagged Hebrew verbs given a non-vocalized verb in base-form and an indication of which pattern the verb follows.

Sequence-to-sequence models (mT5 models) training for Hebrew abstractive summarization.

Morphological Analysis, Disambiguation.

End-to-end pipeline for Hebrew NLP using off the shelf tools, including morphological analysis, tagging, lemmatization, parsing and more.

RoBERTa based language model for Hebrew, present state-of-the-art results on sentiment analysis, named entity recognition and question answering.

State-of-the-art Longformer language model for Hebrew.

Morphological Analysis, disambiguation and dependency Parser. Morphological Analyzer relies on the BGU Lexicon. Demo: https://nlp.biu.ac.il/~rtsarfaty/onlp/hebrew/

An open-source effort to make Hebrew properly searchable by various IR software libraries. Includes Hebrew Analyzer for Lucene.

Tool for Automatic and semi-automatic Nikud for Hebrew texts. Avi Shmidman, Shaltiel Shmidman, Moshe Koppel, and Yoav Goldberg. 2020. Nakdan: Professional Hebrew diacritizer. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics: System Demonstrations, pages 197–203, Online. Association for Computational Linguistics.

Free Hebrew linguistic project including spell checker and morphological analyzer. https://github.com/eranroz/HspellPy/ - Python wrapper for Hspell.

Wikiproject for correcting grammar mistakes. (Heuristic) positive annotations can be derived from query <https://quarry.wmflabs.org/query/21957

Hebrew Diacritizer. Elazar Gershuni and Yuval Pinter: Restoring Hebrew Diacritics Without a Dictionary. Demo in Replicate: https://replicate.com/elazarg/nakdimon/ Paper: https://arxiv.org/abs/2105.05209/ code: https://github.com/elazarg/nakdimon/ data: https://github.com/elazarg/hebrew-diacritize/

Morris Alper's open-source tool for adding vowel signs (Nikud) to Hebrew text, uses no rule-based logic, built with a CANINE transformer network. An interactive demo is available at https://huggingface.co/spaces/malper/unikud Blog post: https://towardsdatascience.com/unikud-adding-vowels-to-hebrew-text-with-deep-learning-powered-by-dagshub-56d238e22d3f

A program to convert Hebrew text files (without Nikud) to text files with the correct Nikud. Developed by Adi Oz and Vered Shani

a BERT model for Hebrew legal and legislative domains. It is intended to improve the legal NLP research and tools development in Hebrew. Two versions: (1) a fine-tuned model of HeBERT applied on legal and legislative documents, and (2) uses HeBERT's architecture guidlines to train a BERT model from scratch.

This code and dataset provide an established benchmark for neural sentiment analysis for Modern Hebrew.

a large pre-trained language model for Modern Hebrew, publicly available, pre-training on Oscar, Texts of Hebrew tweets, all of Hebrew Wikipedia, published by the OnlpLab team.

HeBERT is a Hebrew pretrained language model for Polarity Analysis and Emotion Recognition, published by Dr. Inbal Yahav Shenberger and Avichay Chriqui. It is based on Google's BERT architecture and it is BERT-Base config. HeBert was trained on three dataset: OSCAR, A Hebrew dump of Wikipedia, Emotion User Generated Content (UGC) data that was collected for the purpose of this study. The model was evaluated on downstream tasks: HebEMO - emotion recognition model, and sentiment analysis.

a Hebrew pre-trained language model, trained on the same dataset as the previous SOTA Hebrew PLM AlephBERT, consisting of approximately 2 billion words of text but with a substantially increased vocabulary of 128,000 word pieces. Published as a collaboration of the OnlpLab team and Dicta.

Easy-to-use Python interface for Hebrew clinical psychology text analysis. Useful for various psychology applications such as detecting emotional state, well being, relationship quality in conversation, identifying topics (e.g., family, work) and many more.

a pre-trained BERT model for modern Hebrew, with the masked-language-modeling objective.

A fine-tuned model for prefix segmentation task.

A fine-tuned model for morphological tagging task.

The package is based on Heinrich's java implementation of collapsed Gibbs sampling with an extra variable to model the generative nature of lemmas in Hebrew.