In this paper, we present a semantic parser using a knowledgebase. Instead of relying on filtering the concepts extracted from the knowledgebase, we use all the concepts to create the parser. A simple search is conducted on ConceptNet for the words in the input sentence. In this paper, two proposed techniques are used to extract concepts from the ConceptNet 5. The reason for proposing two techniques in this paper is to address the issue of removing the supervision and training process. The first approach extracts all concepts from ConceptNet 5 for each input word. The extracted concepts are used to search again in ConceptNet 5, which creates multiple levels of search results. This deep concept structure creates a multi-level search to create the semantic parse result. The second approach follows the same first step of extracting concepts using the input text. However, the extracted concepts are passed through a relationship check and then used for the second level search. Concepts are drawn from 2 levels of searching in ConceptNet. The extracted concepts are used to create the parser. Furthermore, we use the initial concepts extracted to search again in ConceptNet. The parser we created is tested on Free917, Stanford Sentiment dataset and the WebQ. We achieve recall of 93.82%, 94.91% for Stanford Sentiment dataset, accuracy of 77.1%, 79.2% for Free917 and 26.5%, 38.2% for WebQ respectively for the two approaches. This shows state-of-the-art results compared to other methods for each datasets.
展开▼
