TextCNN
filters are learnable n-grams
the
cat
sat
on
mat
filter w=2
filter w=3
filter w=4
max-pool + concat
fixed-size vector
LSTM
hidden state carries info forward
h1
h2
h3
h4
h5
the
cat
sat
on
mat
pool h1..h5 → classifier
gates decide what to keep / forget
still fails past ~200 steps
sequential = no parallelism
→ attention came next