Semantic Data Set Construction from Human Clustering and Spatial Arrangement

Abstract Research into representation learning models of lexical semantics usually utilizes some form intrinsic evaluation to ensure that the learned representations reflect human semantic judgments. Lexical similarity estimation is a widely used method, but efforts have typically focused on pairwise judgments words in isolation, or are limited specific contexts and stimuli. There limitations with these approaches either do not provide any context for judgments, thereby ignore ambiguity, very sentential cannot then be generate larger resource. Furthermore, between more than two items considered. We full description analysis our recently proposed methodology large-scale data set construction produces classification large sample verbs first phase, as well multi-way made within resultant classes second phase. The uses spatial multi-arrangement approach field cognitive neuroscience capturing visual adapted this method handle polysemous linguistic stimuli much samples previous work. specifically target verbs, can equally applied other parts speech. perform cluster from phase demonstrate how might useful comprehensive verb also analyze information captured by discuss potential spatially induced better notions word similarity. fine-grained analyses tasks clustering In particular, we find stronger static embedding methods still outperform emerging recent pre-training methods, both word-level clustering. Moreover, thanks set’s vast coverage, able compare benefits specializing vector particular type external knowledge evaluating FrameNet- VerbNet-retrofitted domains such “Heat” “Motion.”