Grounding proposition stores for question answering over linked data

Grounding natural language utterances into semantic representations is crucial for tasks such as question answering and knowledge base population. However, the importance of the lexicons that are central to this mapping remains unmeasured because question answering systems are evaluated as end-to-end systems. This article proposes a methodology to enable a standalone evaluation of grounding natural language propositions into semantic relations by fixing all the components of a question answering system other than the lexicon itself. Thus, we can explore different configurations trying to conclude which are the ones that contribute better to improve overall system performance. Our experiments show that grounding accounts with close to 80% of the system performance without training, whereas training supposes a relative improvement of 7.6%. Finally we show how lexical expansion using external linguistic resources can consistently improve the results from 0.8% up to 2.5%.