DadmaTools is a repository for Natural Language Processing resources for the Persian Language. The aim is to make it easier and more applicable to practitioners in the industry to use Persian NLP, and hence this project is licensed to allow commercial use. The project features code examples on how to use the models in popular NLP frameworks such as spaCy and Transformers, as well as Deep Learning frameworks such as PyTorch. Furthermore, DadmaTools support common Persian embedding and Persian datasets. for more details about how to use this tool read the instruction below.
Contents:
To get started using DadmaTools in your python project, simply install via the pip package. Note that installing the default pip package will not install all NLP libraries because we want you to have the freedom to limit the dependency on what you use. Instead, we provide you with an installation option if you want to install all the required dependencies.
To get started using DadmaTools, simply install the project with pip:
pip install dadmatools Note that the default installation of DadmaTools does install other NLP libraries such as SpaCy and supar.
You can check the requirements.txt file to see what version the packages has been tested with.
Alternatively you can install the latest version from github using:
pip install git+https://github.com/Dadmatech/dadmatools.gitNatural Language Processing is an active area of research, and it consists of many different tasks. The DadmaTools repository provides an overview of Persian models for some of the most basic NLP tasks (and is continuously evolving).
Here is the list of NLP tasks we currently cover in the repository. These NLP tasks are defined as pipelines. Therefore, a pipeline list must be created and passed through the model. This will allow the user to choose the only task needed without loading others. Each task has its abbreviation as follows:
- Named Entity Recognition:
ner - Part of speech tagging:
pos - Dependency parsing:
dep - Constituency parsing:
cons - Kasreh Ezafe Detection:
kasreh - Chunking:
chunk - Lemmatizing:
lem - Tokenizing:
tok - Spellchecker:
spellchecker - Normalizing
- informal2formal:
itf - Sentiment analysis:
sent
Note that the normalizer can be used outside of the pipeline as there are several configs (the default config is in the pipeline with the name of def-norm). Note that if no pipeline is passed to the model, the tokenizer will be loaded as default.
cleaning text and unify characters.
Note: None means no action!
from dadmatools.normalizer import Normalizer
normalizer = Normalizer(
full_cleaning=False,
unify_chars=True,
refine_punc_spacing=True,
remove_extra_space=True,
remove_puncs=False,
remove_html=False,
remove_stop_word=False,
replace_email_with="<EMAIL>",
replace_number_with=None,
replace_url_with="",
replace_mobile_number_with=None,
replace_emoji_with=None,
replace_home_number_with=None
)
text = """
<p>
دادماتولز اولین نسخش سال ۱۴۰۰ منتشر شده.
امیدواریم که این تولز بتونه کار با متن رو براتون شیرین‌تر و راحت‌تر کنه
لطفا با ایمیل [email protected] با ما در ارتباط باشید
آدرس گیت‌هاب هم که خب معرف حضور مبارک هست:
https://github.com/Dadmatech/DadmaTools
</p>
"""
normalized_text = normalizer.normalize(text)
# <p> دادماتولز اولین نسخش سال 1400 منتشر شده. امیدواریم که این تولز بتونه کار با متن رو براتون شیرین‌تر و راحت‌تر کنه لطفا با ایمیل <EMAIL> با ما در ارتباط باشید آدرس گیت‌هاب هم که خب معرف حضور مبارک هست: </p>
# full cleaning
normalizer = Normalizer(full_cleaning=True)
normalized_text = normalizer.normalize(text)
# دادماتولز نسخش سال منتشر تولز بتونه کار متن براتون شیرین‌تر راحت‌تر کنه ایمیل ارتباط آدرس گیت‌هاب معرف حضور مبارکContaining Tokenizer, Lemmatizer, POS Tagger, Dependancy Parser, Constituency Parser, Kasreh, Spellcheker, Infromal To Formal, Name Entity Recognation.
import dadmatools.pipeline.language as language
# here lemmatizer and pos tagger will be loaded
# as tokenizer is the default tool, it will be loaded as well even without calling
pips = 'tok, lem, pos, dep, chunk, cons, spellchecker, kasreh, itf, ner, sent'
nlp = language.Pipeline(pips)
# doc is an SpaCy object
doc = nlp('کشور بزرگ ایران توانسته در طی سال‌ها اغشار مختلفی از قومیت‌های گوناگون رو به خوبی تو خودش جا بده')doc object has different extensions. First, there are sentences in doc which is the list of the list of Token. Each Token also has its own extensions. Note that we defined our own extension as well in DadmaTools. If any pipeline related to the specific extensions is not called, that extension will have no value.
To better see the results which you can use this code:
print(doc){'spellchecker': {'orginal': 'کشور بزرگ ایران توانسته در طی سال\u200cها اغشار مختلفی از قومیت\u200cهای گوناگون رو به خوبی تو خودش جا بده', 'corrected': 'کشور بزرگ ایران توانسته در طی سال\u200cها اقشار مختلفی از قومیت\u200cهای گوناگون رو به خوبی تو خودش جا بده', 'checked_words': [('اغشار', 'اقشار')]}, 'itf': ' کشور بزرگ ایران توانسته در طی سال\u200cها اغشار مختلفی از قومیت های گوناگون را به خوبی در خودش جا بده', 'sentences': [{'id': 1, 'tokens': [{'id': 1, 'text': 'کشور', 'upos': 'NOUN', 'xpos': 'N_PL', 'feats': 'Number=Plur|Person=2|Polarity=Neg|Tense=Pres', 'head': 14, 'deprel': 'nsubj', 'lemma': 'کشور', 'ner': 'O', 'kasreh': 'S-kasreh'}, {'id': 2, 'text': 'بزرگ', 'upos': 'ADJ', 'xpos': 'ADJ', 'feats': 'Degree=Pos', 'head': 8, 'deprel': 'amod', 'lemma': 'بزرگ', 'ner': 'O', 'kasreh': 'S-kasreh'}, {'id': 3, 'text': 'ایران', 'upos': 'SCONJ', 'xpos': 'CON', 'feats': 'Number=Plur|Person=3|PronType=Prs', 'head': 2, 'deprel': 'nmod:poss', 'lemma': 'ایران', 'ner': 'S-loc', 'kasreh': 'O'}, {'id': 4, 'text': 'توانسته', 'upos': 'VERB', 'xpos': 'V_PP', 'feats': 'Number=Sing|Person=3|VerbForm=Part', 'head': 14, 'deprel': 'aux', 'lemma': 'توانست#توان', 'ner': 'O', 'kasreh': 'O'}, {'id': 5, 'text': 'در', 'upos': 'ADP', 'xpos': 'P', 'head': 14, 'deprel': 'case', 'lemma': 'در', 'ner': 'O', 'kasreh': 'O'}, {'id': 6, 'text': 'طی', 'upos': 'ADP', 'xpos': 'P', 'head': 5, 'deprel': 'fixed', 'lemma': 'طی', 'ner': 'O', 'kasreh': 'S-kasreh'}, {'id': 7, 'text': 'سال\u200cها', 'upos': 'AUX', 'xpos': 'V_PRS', 'feats': 'Number=Sing|Person=3|Tense=Pres', 'head': 14, 'deprel': 'fixed', 'lemma': 'سال', 'ner': 'O', 'kasreh': 'O'}, {'id': 8, 'text': 'اغشار', 'upos': 'NOUN', 'xpos': 'N_PL', 'feats': 'Number=Plur', 'head': 19, 'deprel': 'nsubj', 'lemma': 'اغشار', 'ner': 'O', 'kasreh': 'S-kasreh'}, {'id': 9, 'text': 'مختلفی', 'upos': 'ADJ', 'xpos': 'ADJ', 'feats': 'Degree=Pos', 'head': 8, 'deprel': 'amod', 'lemma': 'مختلفی', 'ner': 'O', 'kasreh': 'O'}, {'id': 10, 'text': 'از', 'upos': 'ADP', 'xpos': 'P', 'head': 15, 'deprel': 'case', 'lemma': 'از', 'ner': 'O', 'kasreh': 'O'}, {'id': 11, 'text': 'قومیت\u200cهای', 'upos': 'NOUN', 'xpos': 'N_PL', 'feats': 'Number=Plur', 'head': 8, 'deprel': 'nmod:poss', 'lemma': 'قومیت', 'ner': 'O', 'kasreh': 'S-kasreh'}, {'id': 12, 'text': 'گوناگون', 'upos': 'ADJ', 'xpos': 'ADJ', 'feats': 'Degree=Pos', 'head': 11, 'deprel': 'amod', 'lemma': 'گوناگون', 'ner': 'O', 'kasreh': 'O'}, {'id': 13, 'text': 'رو', 'upos': 'PART', 'xpos': 'CLITIC', 'head': 8, 'deprel': 'case', 'lemma': 'رو', 'ner': 'O', 'kasreh': 'O'}, {'id': 14, 'text': 'به', 'upos': 'ADP', 'xpos': 'P', 'head': 19, 'deprel': 'case', 'lemma': 'به', 'ner': 'O', 'kasreh': 'O'}, {'id': 15, 'text': 'خوبی', 'upos': 'ADJ', 'xpos': 'ADJ', 'feats': 'Degree=Pos', 'head': 14, 'deprel': 'advcl', 'lemma': 'خوب', 'ner': 'O', 'kasreh': 'O'}, {'id': 16, 'text': 'تو', 'upos': 'ADP', 'xpos': 'P', 'feats': 'Number=Sing|Person=2|PronType=Prs', 'head': 19, 'deprel': 'case', 'lemma': 'تو', 'ner': 'O', 'kasreh': 'O'}, {'id': 17, 'text': 'خودش', 'upos': 'PRON', 'xpos': 'PRO', 'feats': 'Number=Sing|Person=3|PronType=Prs|Reflex=Yes', 'head': 19, 'deprel': 'obl', 'lemma': 'خودش', 'ner': 'O', 'kasreh': 'O'}, {'id': 18, 'text': 'جا', 'upos': 'VERB', 'xpos': 'PREV', 'feats': 'Number=Sing|Person=3|Tense=Pres', 'head': 19, 'deprel': 'compound:lvc', 'lemma': 'جا', 'ner': 'O', 'kasreh': 'O'}, {'id': 19, 'text': 'بده', 'upos': 'VERB', 'xpos': 'V_SUB', 'feats': 'Mood=Sub', 'head': 0, 'deprel': 'root', 'lemma': 'داد#ده', 'ner': 'O', 'kasreh': 'O'}]}], 'lang': 'persian', 'sentiment': [{'label': 'positive', 'score': 0.7366364598274231}]}We provide an easy-to-use way to load some popular Persian NLP datasets
Here is the list of supported datasets.
| Dataset | Task |
|---|---|
| PersianNER | Named Entity Recognition |
| ARMAN | Named Entity Recognition |
| Peyma | Named Entity Recognition |
| FarsTail | Textual Entailment |
| FaSpell | Spell Checking |
| PersianNews | Text Classification |
| PerUDT | Universal Dependency |
| PnSummary | Text Summarization |
| SnappfoodSentiment | Sentiment Classification |
| TEP | Text Translation(eng-fa) |
| WikipediaCorpus | Corpus |
| PersianTweets | Corpus |
all datasets are iterator and can be used like below:
from dadmatools.datasets import FarsTail
from dadmatools.datasets import SnappfoodSentiment
from dadmatools.datasets import Peyma
from dadmatools.datasets import PerUDT
from dadmatools.datasets import PersianTweets
from dadmatools.datasets import PnSummary
farstail = FarsTail()
#len of dataset
print(len(farstail.train))
#like a generator
print(next(farstail.train))
#dataset details
pn_summary = PnSummary()
print('PnSummary dataset information: ', pn_summary.info)
#loop over dataset
snpfood_sa = SnappfoodSentiment()
for i, item in enumerate(snpfood_sa.test):
print(item['comment'], item['label'])
#get first tokens' lemma of all dev items
perudt = PerUDT()
for token_list in perudt.dev:
print(token_list[0]['lemma'])
#get NER tag of first Peyma's data
peyma = Peyma()
print(next(peyma.data)[0]['tag'])
#corpus
tweets = PersianTweets()
print('tweets count : ', len(tweets.data))
print('sample tweet: ', next(tweets.data))get dataset info:
from dadmatools.datasets import get_all_datasets_info
get_all_datasets_info().keys()
#dict_keys(['Persian-NEWS', 'fa-wiki', 'faspell', 'PnSummary', 'TEP', 'PerUDT', 'FarsTail', 'Peyma', 'snappfoodSentiment', 'Persian-NER', 'Arman', 'PerSent'])
#specify task
get_all_datasets_info(tasks=['NER', 'Sentiment-Analysis'])the output will be:
{"ARMAN": {"description": "ARMAN dataset holds 7,682 sentences with 250,015 sentences tagged over six different classes.\n\nOrganization\nLocation\nFacility\nEvent\nProduct\nPerson",
"filenames": ["train_fold1.txt",
"train_fold2.txt",
"train_fold3.txt",
"test_fold1.txt",
"test_fold2.txt",
"test_fold3.txt"],
"name": "ARMAN",
"size": {"test": 7680, "train": 15361},
"splits": ["train", "test"],
"task": "NER",
"version": "1.0.0"},
"PersianNer": {"description": "source: https://github.com/Text-Mining/Persian-NER",
"filenames": ["Persian-NER-part1.txt",
"Persian-NER-part2.txt",
"Persian-NER-part3.txt",
"Persian-NER-part4.txt",
"Persian-NER-part5.txt"],
"name": "PersianNer",
"size": 976599,
"splits": [],
"task": "NER",
"version": "1.0.0"},
"Peyma": {"description": "source: http://nsurl.org/2019-2/tasks/task-7-named-entity-recognition-ner-for-farsi/",
"filenames": ["peyma/600K", "peyma/300K"],
"name": "Peyma",
"size": 10016,
"splits": [],
"task": "NER",
"version": "1.0.0"},
"snappfoodSentiment": {"description": "source: https://huggingface.co/HooshvareLab/bert-fa-base-uncased-sentiment-snappfood",
"filenames": ["snappfood/train.csv",
"snappfood/test.csv",
"snappfood/dev.csv"],
"name": "snappfoodSentiment",
"size": {"dev": 6274, "test": 6972, "train": 56516},
"splits": ["train", "test", "dev"],
"task": "Sentiment-Analysis",
"version": "1.0.0"}}To start using embedding please install fasttext:
pip install fasttext
download, load and use some pre-trained Persian word embeddings.
dadmatools supports all glove, fasttext, and word2vec formats.
from dadmatools.embeddings import get_embedding, get_all_embeddings_info, get_embedding_info
from pprint import pprint
pprint(get_all_embeddings_info())
#get embedding information of specific embedding
embedding_info = get_embedding_info('glove-wiki')
#### load embedding ####
word_embedding = get_embedding('glove-wiki')
#get vector of the word
print(word_embedding['سلام'])
#vocab
vocab = word_embedding.get_vocab()
### some useful functions ###
print(word_embedding.top_nearest("زمستان", 10))
print(word_embedding.similarity('کتب', 'کتاب'))
print(word_embedding.embedding_text('امروز هوای خوبی بود'))The following word embeddings are currently supported:
| Name | Embedding Algorithm | Corpus |
|---|---|---|
glove-wiki |
glove | Wikipedia |
fasttext-commoncrawl-bin |
fasttext | CommonCrawl |
fasttext-commoncrawl-vec |
fasttext | CommonCrawl |
word2vec-conll |
word2vec | Persian CoNLL17 corpus |
We have compared our pos tagging, dependancy parsing, and lemmatization models to stanza and hazm.
| PerDT (F1 score) | |||
| Toolkit | POS Tagger (UPOS) | Dependancy Parser (UAS/LAS) | Lemmatizer |
| DadmaTools | 97.52% | 95.36% / 92.54% | 99.14% |
| stanza | 97.35% | 93.34% / 91.05% | 98.97% |
| hazm | - | - | 89.01% |
| Seraji (F1 score) | |||
| Toolkit | POS Tagger (UPOS) | Dependancy Parser (UAS/LAS) | Lemmatizer |
| DadmaTools | 97.83% | 92.5% / 89.23% | - |
| stanza | 97.43% | 87.20% / 83.89% | - |
| hazm | - | - | 86.93% |
| Tehran university tree bank (F1 score) | |
| Toolkit | Constituency Parser |
| DadmaTools (without preprocess)) | 82.88% |
| Stanford (with some preprocess on POS tags) | 80.28 |
You can see the codes and the output in colab.
@inproceedings{etezadi-etal-2022-dadmatools,
title = "{D}adma{T}ools: Natural Language Processing Toolkit for {P}ersian Language",
author = "Etezadi, Romina and
Karrabi, Mohammad and
Zare, Najmeh and
Sajadi, Mohamad Bagher and
Pilehvar, Mohammad Taher",
booktitle = "Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies: System Demonstrations",
month = jul,
year = "2022",
address = "Hybrid: Seattle, Washington + Online",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2022.naacl-demo.13",
pages = "124--130",
abstract = "We introduce DadmaTools, an open-source Python Natural Language Processing toolkit for the Persian language. The toolkit is a neural pipeline based on spaCy for several text processing tasks, including normalization, tokenization, lemmatization, part-of-speech, dependency parsing, constituency parsing, chunking, and ezafe detecting. DadmaTools relies on fine-tuning of ParsBERT using the PerDT dataset for most of the tasks. Dataset module and embedding module are included in DadmaTools that support different Persian datasets, embeddings, and commonly used functions for them. Our evaluations show that DadmaTools can attain state-of-the-art performance on multiple NLP tasks. The source code is freely available at https://github.com/Dadmatech/DadmaTools.",
}