Yousif Haroon, Mohsen Rshwan
Abstract: This paper investigates how universal pre-trained embedding can be adapted to handle the diversity of semantic in specific domains. And present a novel adapting method that relies on the view of semantic constraints to adapt an off the shelf word embedding to close fit the view of the semantics of the target domain. The view of semantic constraints are extracted from quite available resources, a text corpus and a domain-specific dictionary. The view of semantic extraction eliminates the need for rare special-purpose semantic resources and the additional effort for locally trained embedding. The method is implemented as a lightweight final tuning process. The results show that our method outperformed the state of the art embedding adapting methods in the task of Community Question Answering (CQA) for the Arabic medical domain.
Keywords: Word Embedding, Embedding Adaptation, Query Expansion, Transfer Learning, NLP