Improved relation span detection in question answering systems over extracted knowledge bases

In recent years, AI studies have been focused on developing question answering systems to deal with automatic answering natural language questions. Knowledge based open domain question answering systems can generate accurate answers to questions posed by users in various fields. However, these systems need further development to scale the domain of answer retrieval systems and question interpretation. Deep learning methods are one of the current approaches in this research area. Existing knowledge-based question answering systems use either manually curated knowledge bases such as Freebase or knowledge bases automatically extracted from unstructured texts such as Reverb, or a combination of both. In the case of open domain question answering systems, limited access to knowledge bases reduces the expandability of the system. Systems that use only curated knowledge bases have high precision with limited coverage; while systems that use extracted knowledge bases have higher coverage with generally lower precision. To improve the precision of question answering over extracted knowledge bases, this paper presents a solution to enhance detection of the relation span in the question, corresponding to the triples of the extracted knowledge base. First, a dataset with 16,675 simple questions is introduced along with answers based on the Reverb triples. Then, a method based on a fine-tuned BERT model for relation span detection in the questions is proposed. The results showed an increase in the precision of the relation span detection, so that the precision reached 99.65%.

Automated summary

Recent AI studies have focused on developing question answering systems for automatic responses to natural language questions. Knowledge-based open domain question answering systems can accurately generate answers to questions in various fields. However, these systems require further development to scale answer retrieval and question interpretation. Deep learning methods are being used in this research area. Existing knowledge-based question answering systems use manually curated knowledge bases or knowledge bases automatically extracted from unstructured texts, or a combination of both. Limited access to knowledge bases in open domain question answering systems limits their expandability. Systems that use curated knowledge bases have high precision but limited coverage, while systems that use extracted knowledge bases have higher coverage but lower precision. To improve precision over extracted knowledge bases, a solution for enhancing relation span detection in questions is proposed in this paper. A dataset with 16,675 simple questions and answers based on Reverb triples is introduced. A method based on a fine-tuned BERT model is proposed for relation span detection in questions, resulting in a precision of 99.65%.