Pointer Network
A network that answers by pointing back at positions in its input rather than generating brand-new words.
Think of it like
Like answering “which of these do you want?” by pointing at items on the table instead of describing them from scratch.
Example
For extractive summarization, a pointer network selects which input sentences to copy verbatim.
How it actually works
Instead of producing tokens from a fixed vocabulary, it uses attention to point at input elements, which is perfect when the answer is a selection or reordering of the input — sorting, routing, copying exact names. Modern LLMs with copy mechanisms and attention absorbed much of this capability, so pointer networks are now more of a specialized or historical tool.
For product teams
A niche design for tasks that are really about selecting from the input.
For engineers
A seq2seq model whose output distribution is attention over input positions.
Related
- Attention — The mechanism it repurposes to point.
Read anything AI without the jargon
Look up any term in plain English, or save terms as you read with the free Chrome extension.
Open DecoderAdd to Chrome