Character and syllable segmentation give different input units for predicting whole words. The paper compares their language-modeling quality and computational cost; the diagram illustrates the representation choice without displaying experimental measurements. Research source
About this paper
Compares syllable-based and character-based word representations for recurrent neural language models.
Syllable segmentation does not improve modeling quality over character segmentation in the reported experiments. The best syllable-aware model nevertheless achieves comparable performance with 18–33% fewer parameters and trains 1.2–2.2 times faster.
Citation
BibTeX
@inproceedings{takhanov2017syllableawarelanguagemodels,
title = {Syllable-aware Neural Language Models: A Failure to Beat Character-aware Ones},
author = {Assylbekov, Zhenisbek and Takhanov, Rustem and Myrzakhmetov, Bagdat and Washington, Jonathan N.},
year = {2017},
booktitle = {EMNLP 2017},
url = {https://aclanthology.org/D17-1199.pdf}
}