Skip to content

Latest commit

 

History

History
34 lines (21 loc) · 2.87 KB

README.md

File metadata and controls

34 lines (21 loc) · 2.87 KB
Raw

PersianNER

Named-Entity Recognition in Persian Language

ArmanPersoNERCorpus

This is the first manually-annotated Persian named-entity (NE) dataset (ISLRN 399-379-640-828-6). We are releasing it only for academic research use.

The dataset includes 250,015 tokens and 7,682 Persian sentences in total. It is available in 3 folds to be used in turn as training and test sets. Each file contains one token, along with its manually annotated named-entity tag, per line. Each sentence is separated with a newline. The NER tags are in IOB format.

According to the instructions provided to the annotators, NEs are categorized into six classes: person, organization (such as banks, ministries, embassies, teams, nationalities, networks and publishers), location (such as cities, villages, rivers, seas, gulfs, deserts and mountains), facility (such as schools, universities, research centers, airports, railways, bridges, roads, harbors, stations, hospitals, parks, zoos and cinemas), product (such as books, newspapers, TV shows, movies, airplanes, ships, cars, theories, laws, agreements and religions), and event (such as wars, earthquakes, national holidays, festivals and conferences); other are the remaining tokens.

Please cite [1,2] if you use this dataset.

Persian Word Embeddings

GloVe300d, word2vec_cbow300d, word2vec_skipgram300d, and hpca300d are four different word embeddings trained on a sizable collation of unannotated Persian text. They contain a comprehensive Persian dictionary of nearly 50K unique words. The length of the embedding vectors is 300. The use of these embeddings is unrestricted.

Please cite [2] if you use any of these word embeddings in your work.

Reported Scores

Paper CoNLL F1 Score*
[1] 65.13%
[2] 77.45%
https://ieeexplore.ieee.org/abstract/document/8661067 76.79%
https://ieeexplore.ieee.org/abstract/document/8700549 84.23%

*The CoNLL F1 score is a strict version of the standard F1 score where a true positive is scored only if all the tokens of a given named entity are classified correctly (including their B- and I- tags). Conversely, every incorrect B- prediction is counted as a false positive. You may find the conlleval Perl script here https://www.clips.uantwerpen.be/conll2000/chunking/conlleval.txt .

Citations

  1. Hanieh Poostchi, Ehsan Zare Borzeshi, Mohammad Abdous, and Massimo Piccardi, "PersoNER: Persian Named-Entity Recognition," The 26th International Conference on Computational Linguistics (COLING 2016), pages 3381–3389, Osaka, Japan, 2016.

  2. Hanieh Poostchi, Ehsan Zare Borzeshi, and Massimo Piccardi, "BiLSTM-CRF for Persian Named-Entity Recognition; ArmanPersoNERCorpus: the First Entity-Annotated Persian Dataset," The 11th Edition of the Language Resources and Evaluation Conference (LREC), Miyazaki, Japan, 7-12 May 2018, ISLRN 399-379-640-828-6, ISLRN 921-509-141-609-6.