Spatial Relation Extraction on AMR Enhancement and Additional Markers

Understanding spatial relations is one of the key areas in natural language processing and artificial intelligence. Previous work only considered a single dimension of sentence token representation, ignoring the semantic structure information contained in the sentence. Abstract Meaning Representation (AMR) graphs can transform natural language text into structured representations with rich semantics. In this work, we introduce AMR graph aggregation to capture the semantic information of sentences and obtain the embedding after fusing the semantic structures by parsing the sentence with a pre-trained AMR parser. Furthermore, we incorporate the syntactic information of the triplets by adding typed text markers. Experimental results on the SpaceEval dataset show that our approach outperforms the strong baselines.