Improving Persian Word Embeddings Using Cross-Lingual Joint Space

In this paper, cross-lingual word embeddings method is introduced to improve quality of monolingual Persian word embeddings. The main idea of the paper is that as Persian is low resource language, a high resource language like English can enhance Persian word embeddings in a cross-lingual space. Therefore, English monolingual word embeddings are used to create a joint space with Persian one. MUSE and VecMap method as the two state of the art approaches are applied to transfer Persian word embeddings to English word embeddings space in a supervised mode. A 5k bilingual English-Persian is utilized as the supervision. In addition, the English Kudkudak evaluation benchmark is customized to assess Persian monolingual word embeddings. The customized benchmark evaluates word embeddings in three tasks, namely categorization, analogy, and word similarity. According to analysis, the cross-lingual transfer can increase monolingual Persian word embeddings without any extra train data. In comparison with MUSE, VecMap can align Persian word embeddings to English in a more effective way. As a result, VecMap outperforms MUSE in enhancing Persian word embeddings. The Persian cross-lingual word embeddings show improvements in categorization and analogy tasks.