Predicate Detection using Word Embeddings for Question Answering over Linked Data

Knowledge-based question-answering system (KBQA) has demonstrated an ability to generate answers of natural language from information stored in a large-scale knowledge base. It has attracted a lot of attentions in the research areas of natural language processing and information retrieval. Generally, it complete the analysis challenge via three steps: identifying named entities, detecting predicates and generate SPARQL queries. In these three steps, predicate detection, a core component of this process, identifies the KB relation(s) a question refers to. To build a predicate detection structure, we identify all possible named entity first, then collect all predicates corresponding to the above entities. What follows is to calculate the similarity between problem and candidate predicates using a Multi-granularity neural network model(MGNN). To find the globally optimal entity-predicate assignment, we use a joint model which is based on the result of entity linking and predicate detection process rather than considering the local predictions (i.e. most possible entity or predicate) as the final result.