--- annotations_creators: - SLPL language_creators: - SLPL language: - fa license: - mit multilinguality: - monolingual size_categories: - 200M ## Dataset Creation ### Curation Rationale Due to the lack of a huge amount of text data in lower resource languages - like Farsi - researchers working on these languages were always finding it hard to start to fine-tune such models. This phenomenon can lead to a situation in which the golden opportunity for fine-tuning models is just in hands of a few companies or countries which contributes to the weakening the open science. The last biggest cleaned merged textual corpus in Farsi is a 70GB cleaned text corpus from a compilation of 8 big data sets that have been cleaned and can be downloaded directly. Our solution to the discussed issues is called naab. It provides **126GB** (including more than **224 million** sequences and nearly **15 billion** words) as the training corpus and **2.3GB** (including nearly **11 million** sequences and nearly **300 million** words) as the test corpus. ### Source Data The textual corpora that we used as our source data are illustrated in the figure below. It contains 5 corpora which are linked in the coming sections.
#### Persian NLP [This](https://github.com/persiannlp/persian-raw-text) corpus includes eight corpora that are sorted based on their volume as below: - [Common Crawl](https://commoncrawl.org/): 65GB ([link](https://storage.googleapis.com/danielk-files/farsi-text/merged_files/commoncrawl_fa_merged.txt)) - [MirasText](https://github.com/miras-tech/MirasText): 12G - [W2C – Web to Corpus](https://lindat.mff.cuni.cz/repository/xmlui/handle/11858/00-097C-0000-0022-6133-9): 1GB ([link](https://storage.googleapis.com/danielk-files/farsi-text/merged_files/w2c_merged.txt)) - Persian Wikipedia (March 2020 dump): 787MB ([link](https://storage.googleapis.com/danielk-files/farsi-text/merged_files/fawiki_merged.txt)) - [Leipzig Corpora](https://corpora.uni-leipzig.de/): 424M ([link](https://storage.googleapis.com/danielk-files/farsi-text/merged_files/LeipzigCorpus.txt)) - [VOA corpus](https://jon.dehdari.org/corpora/): 66MB ([link](https://storage.googleapis.com/danielk-files/farsi-text/merged_files/voa_persian_2003_2008_cleaned.txt)) - [Persian poems corpus](https://github.com/amnghd/Persian_poems_corpus): 61MB ([link](https://storage.googleapis.com/danielk-files/farsi-text/merged_files/poems_merged.txt)) - [TEP: Tehran English-Persian parallel corpus](http://opus.nlpl.eu/TEP.php): 33MB ([link](https://storage.googleapis.com/danielk-files/farsi-text/merged_files/TEP_fa.txt)) #### AGP This corpus was a formerly private corpus for ASR Gooyesh Pardaz which is now published for all users by this project. This corpus contains more than 140 million paragraphs summed up in 23GB (after cleaning). This corpus is a mixture of both formal and informal paragraphs that are crawled from different websites and/or social media. #### OSCAR-fa [OSCAR](https://oscar-corpus.com/) or Open Super-large Crawled ALMAnaCH coRpus is a huge multilingual corpus obtained by language classification and filtering of the Common Crawl corpus using the go classy architecture. Data is distributed by language in both original and deduplicated form. We used the unshuffled-deduplicated-fa from this corpus, after cleaning there were about 36GB remaining. #### Telegram Telegram, a cloud-based instant messaging service, is a widely used application in Iran. Following this hypothesis, we prepared a list of Telegram channels in Farsi covering various topics including sports, daily news, jokes, movies and entertainment, etc. The text data extracted from mentioned channels mainly contains informal data. #### LSCP [The Large Scale Colloquial Persian Language Understanding dataset](https://iasbs.ac.ir/~ansari/lscp/) has 120M sentences from 27M casual Persian sentences with its derivation tree, part-of-speech tags, sentiment polarity, and translations in English, German, Czech, Italian, and Hindi. However, we just used the Farsi part of it and after cleaning we had 2.3GB of it remaining. Since the dataset is casual, it may help our corpus have more informal sentences although its proportion to formal paragraphs is not comparable. #### Initial Data Collection and Normalization The data collection process was separated into two parts. In the first part, we searched for existing corpora. After downloading these corpora we started to crawl data from some social networks. Then thanks to [ASR Gooyesh Pardaz](https://asr-gooyesh.com/en/) we were provided with enough textual data to start the naab journey. We used a preprocessor based on some stream-based Linux kernel commands so that this process can be less time/memory-consuming. The code is provided [here](https://github.com/Sharif-SLPL/t5-fa/tree/main/preprocess). ### Personal and Sensitive Information State whether the dataset uses identity categories and, if so, how the information is used. Describe where this information comes from (i.e. self-reporting, collecting from profiles, inferring, etc.). See [Larson 2017](https://www.aclweb.org/anthology/W17-1601.pdf) for using identity categories as a variables, particularly gender. State whether the data is linked to individuals and whether those individuals can be identified in the dataset, either directly or indirectly (i.e., in combination with other data). State whether the dataset contains other data that might be considered sensitive (e.g., data that reveals racial or ethnic origins, sexual orientations, religious beliefs, political opinions or union memberships, or locations; financial or health data; biometric or genetic data; forms of government identification, such as social security numbers; criminal history). If efforts were made to anonymize the data, describe the anonymization process. ## Considerations for Using the Data ### Social Impact of Dataset Farsi is a language used by millions of people, for thousands of years; therefore, there exists numerous resources for this language. However, no-one has ever published a big enough easy to use corpus of textual data. Our dataset eases the path of pre-training and fine-tuning Farsi Language Models (LMs) in self-supervised manner which can lead to better tools for retention and development of Farsi. As a matter of fact, the informal portion of naab contains various dialects including, Turkish, Luri, etc. which are under-represented languages. Although the amount of data is comparably small, but it can be helpful in training a multi-lingual Tokenizer for Farsi variations. As mentioned before, some parts of our dataset are crawled from social media which in result means it contains ethnic, religious, and gender biases. ### Discussion of Biases During Exploratory Data Analysis (EDA), we found samples of data including biased opinions about race, religion, and gender. Based on the result we saw in our samples, only a small portion of informal data can be considered biased. Therefore, we anticipate that it won’t affect the trained language model on this data. Furthermore, we decided to keep this small part of data as it may become helpful in training models for classifying harmful and hateful texts. ### Other Known Limitations If studies of the datasets have outlined other limitations of the dataset, such as annotation artifacts, please outline and cite them here. ## Additional Information ### Dataset Curators List the people involved in collecting the dataset and their affiliation(s). If funding information is known, include it here. ### Licensing Information Provide the license and link to the license webpage if available. ### Citation Information Provide the [BibTex](http://www.bibtex.org/)-formatted reference for the dataset. For example: ``` @article{article_id, author = {Author List}, title = {Dataset Paper Title}, journal = {Publication Venue}, year = {2525} } ``` If the dataset has a [DOI](https://www.doi.org/), please provide it here. ### Contributions Thanks to [@sadrasabouri](https://github.com/sadrasabouri) and [@elnazrahmati](https://github.com/elnazrahmati) for adding this dataset.