A novel unsupervised method for named-entity identification in resource-poor languages using bilingual corpus

We propose a new unsupervised method to identify Named Entities (NE) in resource-poor languages. The idea is to transfer the knowledge of NEs from a resource-rich language to a resource-poor one by using a bilingual parallel corpus of this language pair. After extracting all NE pair candidates and filtering these candidates (includes lexical and contextual filters) to obtain a high precision seed of NEs, a graph is created for each language using these seeds. This graph is used for bootstrapping of the primary seeds. Based on output of the graph, a classifier is trained to identify NEs in the resource-poor language. In this paper, Farsi and English are selected as representatives for resource-poor and resource-rich languages, respectively. Because Farsi is a non-Latin language, we present a new distance function called M-distance to compute edit distance between Latin and Farsi scripts. Finally, we released a Farsi NE identifier (without using specific features of Farsi) for the first time with F1 score of 0.74.