Search for: All records

Like speech, signs are composed of discrete, recombinable features called phonemes. Prior work shows that models which can recognize phonemes are better at sign recognition, motivating deeper exploration into strategies for modeling sign language phonemes. In this work, we learn graph convolution networks to recognize the sixteen phoneme “types” found in ASL-LEX2.0. Specifically, we explore how learning strategies like multi-task and curriculum learning can leverage mutually useful information between phoneme types to facilitate the remodeling of sign language phonemes. Results on the Sem-Lex Benchmark show that curriculum learning yields an average accuracy of 87% across all phoneme types, outperforming fine-tuning and multi-task strategies for most phonemetypes. 

abstract A growing body of research shows that both signed and spoken languages display regular patterns of iconicity in their vocabularies. We compared iconicity in the lexicons of American Sign Language (ASL) and English by combining previously collected ratings of ASL signs (Caselli, Sevcikova Sehyr, Cohen-Goldberg, & Emmorey, 2017) and English words (Winter, Perlman, Perry, & Lupyan, 2017) with the use of data-driven semantic vectors derived from English. Our analyses show that models of spoken language lexical semantics drawn from large text corpora can be useful for predicting the iconicity of signs as well as words. Compared to English, ASL has a greater number of regions of semantic space with concentrations of highly iconic vocabulary. There was an overall negative relationship between semantic density and the iconicity of both English words and ASL signs. This negative relationship disappeared for highly iconic signs, suggesting that iconic forms may be more easily discriminable in ASL than in English. Our findings contribute to an increasingly detailed picture of how iconicity is distributed across different languages.