Commits

Author Commit Message Labels Comments Date
David McClosky
first-stage/TRAIN/InputTree.C: Allow ROOT as a legal toplevel type
David McClosky
second-stage/programs/wlle/cvlm-lbfgs.cc: Fix weights writing bug This fixes a segfault if your weights file is set to be in a directory that doesn't exist.
David McClosky
python/bllipparser/ModelFetcher.py: Add DK's WSJ-PTB3 unified parsing model Various docs updated to prefer this over my less standard WSJ models.
David McClosky
MODELS.rst: New docs on which parsing model to use README.rst, README-python.rst: point to MODELS.rst
David McClosky
second-stage/README-retrain.rst: $NBEST is really $NPARSES Thanks DK for catching this!
David McClosky
second-stage/README-retrain.rst: Update training estimate
David McClosky
python/bllipparser/ModelFetcher.py: WSJ+PubMed -> GENIA+PubMed This model file wasn't named correctly. The old download URL will still work, though.
David McClosky
setup.py: Parser SWIG depends on SimpleAPI.*
David McClosky
python/bllipparser/RerankingParser.py: parse_constrained does a 2nd pass This is similar to what Tree.log_prob() does under the covers. There's also a bug fix for NBestList.__str__ which didn't work correctly in cases where there are no parses in the list when there's no reranker.
David McClosky
first-stage/PARSE/SimpleAPI.h: Expose SpanConstraints.minSizeForParsing This is needed to allow RerankingParser.parse_constrained to try with multiple passes.
David McClosky
first-stage/PARSE/SimpleAPI.C: Remove old decodeOnly option This option was used in a previous (uncommitted) implementation of parser constraints.
David McClosky
python/bllipparser/RerankingParser.py: allow ExtPos short circuiting
David McClosky
python/bllipparser/RerankerFeatureCorpus.py: More reranker tools This adds many new methods if you want to build a reranker using a pure Python feature extractor. Tools help you count features to build feature indices, prune them in a similar way as the real reranker (based on how often they distinguish "winners" from "losers"), perform the mapping, relativizing feature counts, computing f-scores for parses and oracle parses, etc.
David McClosky
python/bllipparser/RerankingParser.py: collapsed parse_tagged() It's now just a special case of parse_constrained()
David McClosky
python/bllipparser/RerankingParser.py: handle parse failures properly Also added a test for this.
David McClosky
Makefile: real-clean is even cleaner
David McClosky
first-stage/PARSE/parseIt.C: finally adding a -h option "This has gone on just long enough!"
David McClosky
first-stage/PARSE/parseIt.C: remove unused variable
David McClosky
first-stage/PARSE/SimpleAPI.C: Remove unused variable
David McClosky
first-stage/PARSE/SimpleAPI.[Ch]: rename apply -> applyToChart SWIG doesn't like the name "apply". To err on the side of safety, we're changing it instead of just renaming it in Python's wrapper.
David McClosky
first-stage/PARSE/swig/wrapper.i: Move some templates later
David McClosky
.travis.yml: Print out SWIG version
David McClosky
Added constrained parsing option, various other improvements New constrained parsing mechanism: This allows for partially or completely constrained parses (each constraint takes the form of a span and a set of possible labels allowed to cover that span). There was a previous mechanism for this, but it had many quirks which made it unintuitive to use. The new mechanism requires the exact parsing constraints to be decided before parsing time. A new class, SpanConstraints, handles the tricky aspects of this so that only the required constraints need to be expressed. The Python parse_constrained() method makes it even easier. SimpleAPI.C: The non-SWIG functions of wrapper.i have been refactored into the new SimpleAPI.C, similar in spirit to simple-api.cc in the second-stage. SWIG generated files now go in the swig/ subdirectories to keep them separate. Details: first-stage/PARSE/SimpleAPI.C: In addition to above refactoring: - main parse() method uses new constrained parsing system - added LabeledSpan and Spa…
David McClosky
first-stage/PARSE/parseIt.C: more types to size_t
David McClosky
first-stage/PARSE/Bchart.[Ch]: Nth is a size_t This simplifies comparisons with other sizes (and reduces warnings).
David McClosky
python/examples/evaluate.py: Example evaluation code
David McClosky
python/examples/tree_fun.py: Whitespace cleanup
David McClosky
python/examples/tree_diff.py: Simple tree differ
David McClosky
first-stage/PARSE/Item.h: Add a comment (!)
David McClosky
CHECKLIST.txt: more on flake8 use
  1. Prev
  2. Next