reads words like a detective reads ransom notes, one letter at a time.
means a convolutional neural network that learns from characters instead of whole words, building meaning up from letters, prefixes, and typos.
from emerged around 2015 as researchers like yoon kim and xiang zhang asked what happens if you skip word embeddings entirely and let cnns slide filters over raw character sequences, catching morphology and misspellings that word-level models miss.
kim et al 2016 — character-aware language model, beat word-level baselines on ptb
zhang and lecun 2015 — char-level cnn for text classification, no words needed at all
elmo character encoder — used charcnn layers before feeding into bilstm, 2018
fasttext subword tricks — related idea, embeddings built from character n-grams