All
›
lab
›
guide
›
examples
›
graph
/lab/guide/examples/graph
0 folders · 5 notes
▦ Grid
▤ List
≣ Compact
Name
Updated
Branch B — LLM paraphrase augmentation (the route that won)
The other spoke of the branching example ( lab/guide/06-forks-and-graph ): same trigger, opposite outcome. Paraphrase each CoNLL sentence with the local LLM, keeping entity spans verbatim; 3x augmented train set. metrics:: TweetNER dev F1 8
by Lab Guide · edited 2026-07-03
Jul 3
☆
Open questions raised by the augmentation work
A forward branch of the branching example ( lab/guide/06-forks-and-graph ): where the winning approach points next. Open: does span-pinned paraphrase augmentation still help for low-resource, non-English NER, where the LLM's paraphrases are
by Lab Guide · edited 2026-07-03
Jul 3
☆
Results log — the evidence both branches point at
The shared evidence note of the branching example ( lab/guide/06-forks-and-graph ). Claims and experiments elsewhere cite it with evidence:: lab/guide/examples/graph/results-log . Baseline: CoNLL-trained BERT-base tagger, evaluated cross-do
by Lab Guide · edited 2026-07-03
Jul 3
☆
Robust NER on noisy text — the hub note
This is the hub of the seeded branching example (see lab/guide/06-forks-and-graph ). Its callouts reference the other notes, which is what makes the /graph (/graph) view a tree instead of a line. We noticed our NER model falls apart on soci
by Lab Guide · edited 2026-07-03
Jul 3
☆
Branch A — rule-based noising (dead end)
One spoke of the branching example: the route that died , recorded so the lesson survives. Part of lab/guide/06-forks-and-graph . Character-level noise + hashtag/emoji injection over CoNLL train. metrics:: TweetNER dev F1 79.1 (baseline 78.
by Lab Guide · edited 2026-07-03
Jul 3
☆