FormNet
Google's structure-aware form extractor — Rich Attention plus graph-built Super-Tokens to recover the layout that serialization destroys.
About
FormNet starts from a practical complaint: turning a form into a token sequence destroys the 2D relationships that make the form readable, and no serialization heuristic fully recovers them. Google's fix has two parts. Rich Attention bakes the spatial relationship between tokens directly into the attention computation, and Super-Tokens aggregate each word's neighborhood through a graph convolutional network before the sequence model ever sees them. Extended Transformer Construction (ETC) layers then handle the long sequences forms produce, with Viterbi decoding on top.
Presented at ACL 2022 (arXiv 2203.08411), it reported state-of-the-art on the CORD, FUNSD, and Payment benchmarks with a more compact model and less pre-training data than competitors — the efficiency angle again.
Like DocFormer, this is a paper and a method rather than a downloadable model: Google published the research but no official checkpoint release could be confirmed, so plan on re-implementation.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
- Author-reported (ACL 2022, arXiv 2203.08411): state-of-the-art on CORD, FUNSD, and Payment benchmarks with a more compact model and less pre-training data (provider-claimed)
- No confirmed official weight release, so no independent replication of the exact model exists
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 6/10
Scored 6/10 (TIER verified, metrics-v1). Evidence: peer-reviewed ACL 2022 paper from Google Cloud AI Research / Google Research (Lee et al.) with a well-specified architecture (Rich Attention, Super-Tokens, ETC) and strong author-reported results on three benchmarks. Strengths: peer-reviewed, influential structure-aware ideas, honest efficiency story. Docked for: (1) no confirmed official weights or maintained code — paper plus recipe; (2) all numbers are author-reported. License: research publication; no released artifact to license.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 6.0/10
- Maintenance · recency, releases, issue load
- 4.0/10
- Openness · license + docs
- 5.0/10
- Manual review · sanity read + red-flag check
- 7.0/10
- · Peer-reviewed ACL 2022 paper from Google Cloud AI Research / Google Research (Lee et al.): Rich Attention, GCN-built Super-Tokens, ETC layers, Viterbi decoding.
- · Author-reported SOTA on CORD, FUNSD, and Payment benchmarks with a compact model and less pre-training data.
- · No confirmed official weight release — a paper plus a recipe, not a downloadable model.