[
https://issues.apache.org/jira/browse/OPENNLP-1889?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=18096636#comment-18096636
]
Kristian Rickert commented on OPENNLP-1889:
-------------------------------------------
Linking the design
> OpenNLP modernization: document-shape pipeline and pure-JVM PoC feature set
> ---------------------------------------------------------------------------
>
> Key: OPENNLP-1889
> URL: https://issues.apache.org/jira/browse/OPENNLP-1889
> Project: OpenNLP
> Issue Type: Epic
> Reporter: Kristian Rickert
> Assignee: Kristian Rickert
> Priority: Major
> Fix For: 3.0.0
>
>
> A proof-of-concept modernization of OpenNLP around an immutable,
> offset-anchored
> {{Document}} model, plus a set of pure-JVM analysis capabilities built on top
> of
> it. It was developed as one integrated branch and is being split into small,
> independently reviewable pieces. The goal is a single-jar, no-Python,
> span-faithful
> NLP stack that we can expose over gRPC as a serious alternative to the popular
> Python packages.
> h2. Rollout policy (applies to every ticket in this epic)
> - Each feature is developed on its own branch and is *not* pushed to apache
> until it
> is ready for collaboration and review. These drafts exist so we can cut
> branches in
> our own environment first. I'd love to share, so just reply and I can make a
> branch for collaboration and open to any changes along the way. I'd love the
> help.
> - *Only the document shape (OPENNLP-1888) is considered necessary for 3.0.*
> Every other ticket here can land after 3.0 without blocking the release.
> - As each feature lands, we add support for it to the gRPC service, so the
> capability
> is exposed over the wire as it becomes available.
> - Every ticket and feature will have Javadoc, unit tests, eval tests, manual
> entries, and samples that are runnable. No exceptions.
> - Any test results will me runnable and repeatable. We will focus most on
> industry standard tests. Java is fast; I have full confidence we will have
> even more impressive results than what you see below.
> - All caveats will be documented (i.e. Java version of Model2Vec does lower
> accuracy - at times too much - for 20x speedup.)
> h2. PoC results (measured this cycle)
> ||Area||Result||Reference / target||
> |Static embeddings (model2vec / potion-base-8M), pure-JVM inference|12.9x
> faster single-thread, ~7x peak throughput multi-thread, at 0.22x memory|vs
> the Python reference; vector parity verified *before* optimizing|
> |Dependency parser (pure-JVM feedforward neural, trained in-process)|86.78
> UAS / 84.61 LAS with gold tags (h400, beam 4)|UD English EWT test|
> |Full all-neural pipeline (our tagger -> our parser)|83.73 UAS / 79.88 LAS at
> 448 tokens/s|Stanza end-to-end 88.90 / 86.77|
> |Feedforward POS tagger|94.68%|classical maxent ceiling 93.75%|
> |CJK segmentation (mecab-format Viterbi lattice, char-trie)|5.1M chars/s,
> ~146k sentences/s; loads 392k IPADIC entries in ~0.7s|real IPADIC; one engine
> covers JA + KO|
> |UD lemmatizer (via the installer)|87.76% lemma accuracy (~22s train, default
> params)|UD English EWT|
> |Normalization / tokenization hot paths (OPENNLP-1878, #1161)|~2x on the
> measured hot paths, non-breaking|already in review|
> h2. Training and data provenance (legality)
> - *We do our own training.* The feedforward parser and tagger are trained
> in-process
> in the JVM from open treebanks; the embedding tables are verified for parity
> in-JVM. *No third-party model binaries are shipped.* We would need to find a
> home for the newly trained models.
> - *Nothing third-party is bundled.* A bring-your-own-data installer takes a
> user-supplied URL, verifies SHA-256 before unpacking, and rejects path
> escapes. This is how mecab dictionaries, UD treebanks, Hunspell dictionaries,
> gazetteers, and Who's On First tables are acquired: the user points at the
> source and thereby accepts its license. (Cleared via LEGAL-732)
> - *Geographic data:* the only bundled gazetteer is Natural Earth populated
> places
> (public domain). GeoNames (CC-BY) and Overture are download-only. Who's On
> First is BYOD-only (a large multi-source license patchwork).
> Census/IRS-derived place profiles are gated behind legal review and never
> bundled: the engine ships, the data does not.
> - *Benchmark data:* UD English EWT is CC-BY-SA and used for measurement
> only; we do not ship models derived from it.
>
> I'm cleaning up the code now - but I'd be glad to put them as branches but
> it's still in PoC and the stacked dependency makes it unrealistic and very
> confusing. Once OPENNLP-1888 is committed, the complexity of the merge order
> goes down and most tickets can land off main.
> h3. The coolest part -
> I try not to hype or get too opinionated, especially in a Jira ticket, but
> I'm thrilled to work with anyone welcome on this. To say the least - Java is
> a great language and initial numbers suggest that we'll gain a lot of
> customers should we land these features. This opportunity is huge.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
