Monday February 18, 2013 at 16h30 in Celestijnenlaan 200A (auditorium 00.225)
Structured Machine Learning for Mapping Natural Language to Spatial Ontologies
by Parisa Kordjamshidi
We extend the task of spatial role labeling to the task of mapping natural language to spatial ontologies based on multiple spatial calculi. We propose a global structured learning framework for this task. In this framework spatial roles and relations in a spatial role labeling (SpRL) layer and, the semantics of the relations in a spatial qualitative labeling (SpQL) layer are extracted in a global learning model. However given the large possible output spaces according to the spatial ontology, the global inference-based structured learning becomes intractable. We analyze various model compositions and decompositions in the framework of structured learning. To achieve a tractable model we propose a decomposed inference model during training and prediction based on the two semantic layers (SpRL and SpQL). Using the proposed decomposition named communicative inference, our model outperforms the pipelining of the two layers as well as other relevant decomposed learning models when evaluated on textual descriptions of CLEF IAPR TC-12 benchmark. Our work is an important step towards automatically describing text with semantic labels that form a structured ontological representation of its content.