Improving A Domain-Specific Sentence Semantic Representation with Non-Relevant Domain Data

Abstract

The semantic vector representation of sentences in specific domain has important uses and plays an important role in question answering and retrieval. However, the semantic representation of sentences obtained based on neural networks requires sufficient training data. And a specific domain may lack data for semantic training, and it is difficult to obtain such data. And there may already be a lot of data for training languages in many other specific domains. This paper attempts to use these semantic training data independent of the target domain to train a semantic model for the target domain to help improve the semantic representation of this domain. Experiments show that using a BERT-based semantic representation model, trained with non-relevant domain data, can effectively improve the semantic vector representation of target specific domain.