BiLSTM-CRF for Persian Named-Entity Recognition ArmanPersoNERCorpus: the First Entity-Annotated Persian Dataset

BiLSTM-CRF for Persian Named-Entity Recognition ArmanPersoNERCorpus: the First Entity-Annotated Persian Dataset Hanieh Poostchia,b, Ehsan Zare Borzeshib, Massimo Piccardia a University of Technology Sydney, b Capital Markets CRC PO Box 123 Broadway NSW 2007, Australia; 55 Harrington Street The Rocks NSW 2000, Australia {hpoostchi,ezborzeshi}@cmcrc.com, massimo.piccardi@uts.edu.au Abstract Named-entity recognition (NER) can still be regarded as work in progress for a number of Asian languages due to the scarcity of annotated corpora. For this reason, with this paper we publicly release an entity-annotated Persian dataset and we present a performing approach for Persian NER based on a deep learning architecture. In addition to the entity-annotated dataset, we release a number of word embeddings (including GloVe, skip-gram, CBOW and Hellinger PCA) trained on a sizable collation of Persian text. The combination of the deep learning architecture (a BiLSTM-CRF) and the pre-trained word embeddings has allowed us to achieve a 77.45% CoNLL F1 score, a result that is more than 12 percentage points higher than the best previous result and interesting in absolute terms. Keywords: Named-entity recognition, recurrent neural networks, BiLSTM-CRF, Persian language, low-resource languages. 1. Introduction Named-entity recognition (NER) is a natural language processing component that aims to identify all the “named entities” (NEs) such as names of people, locations, organi- sations and numerical expressions in an unstructured text. This information can be useful in its own right or facilitate higher-level NLP tasks such as text summarization and machine translation. To date, NER research has mostly focussed on languages with a high number of digitally annotated resources such as English and German (Tjong Kim Sang and De Meulder, 2003) and Spanish and Dutch (Tjong Kim Sang, 2002). The main reason why many other languages, including many from the Asian region, have received less attention is the significant scarcity of public, annotated digital resources. Amongst those, the Persian language is spoken by more than 110 million speakers world-wide and has more than 570K articles on Wikipedia. However, it has been rarely studied for NER (Khormuji and Bazrafkan, 2014) or even just text processing (Shamsfard, 2011). Although language-agnostic NER systems such as Polyglot-NER (Al-Rfou et al., 2015) exist, their per- formance is generally not competitive in comparison to language-specific NER. For this reason, in our previous work (Poostchi et al., 2016) we developed a dedicated NER system for Persian1. Its development was supported by two datasets: a) a sizable unannotated dataset of Persian sentences for training word embeddings, and b) an entity- annotated dataset for training named-entity classifiers. This paper makes three distinct contributions: 1) it offi- cially releases the entity-annotated dataset with an ISLRN2 that should make its utilisation easier; 2) it releases four different word embeddings trained on the unannotated 1Particularly, Western/Iranian Persian which is also known as Farsi. 2ISLRN: 399-379-640-828-6 resources for a comprehensive Persian dictionary of nearly 50K unique words, also available via an ISLRN3; 3) it proposes a deep learning Persian NER based on a state-of-the-art architecture, the BiLSTM-CRF (Huang et al., 2015; Lample et al., 2016). Thanks to this architecture …

Content truncated to 500 words