Complementing language embeddings with knowledge bases for specific domains

Language embeddings are a promising approach for handling natural language expressions. Current embeddings encompass a large language corpus, and need to be retrained to deal with specific sub-domains. On the other hand, these embeddings often disregard even basic domain knowledge, making them specially fragile when handling technical, specific, knowledge domains, and requiring costly retraining. To alleviate this issue, we propose a combined approach where the embedding is seen as a model of a logical knowledge base. Through a continuous learning approach, the embedding improves its satisfaction of the knowledge base, and in turn produces better training examples by labelling previously unseen text. In this position paper we describe the general framework for this continuous learning, along with its main features.