| Methods for estimating people's conceptual knowledge have the potential to be very useful to theoretical research on conceptual semantics. Traditionally, feature-based conceptual representations have been estimated using property norm data; however, computational techniques have the potential to build such representations automatically. The automatic acquisition of feature-based conceptual representations from corpora is a challenging task, given the unconstrained nature of what can constitute a semantic feature. Existing computational methods typically do not target the full range of concept-relation-feature triples occurring in human generated norms (e.g. tiger have stripes) but rather focus on concept-feature tuples (e.g. tiger – stripes) or triples involving specific relations only. We investigate the large-scale extraction of concept-relation-feature triples and the usefulness of encyclopedic, syntactic and semantic information in guiding the extraction process. Our method extracts candidate triples (e.g. tiger have stripes, flute produce sound) from parsed corpus data and ranks them on the basis of semantic information. Our investigation shows the usefulness of external knowledge in guiding feature extraction and highlights issues of methodology and evaluation which need to be addressed in developing models for this task. |
Récupérer au format
