[
https://issues.apache.org/jira/browse/OPENNLP-1991?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=18126312#comment-18126312
]
ASF GitHub Bot commented on OPENNLP-1991:
-----------------------------------------
krickert commented on code in PR #1323:
URL: https://github.com/apache/opennlp/pull/1323#discussion_r4247904090
##########
opennlp-core/opennlp-formats/src/test/java/opennlp/tools/formats/conllu/ConlluDependencyParserUsageTest.java:
##########
@@ -211,4 +221,49 @@ void testPersistedModelParsesLikeTheOriginal(@TempDir Path
dir) throws IOExcepti
reloaded.parse(new String[] {"she", "eats", "fish"},
new String[] {"PRON", "VERB", "NOUN"}));
}
+
+ @Test
+ void testFeedforwardTrainSaveReloadAndParse(@TempDir Path dir) throws
IOException {
+ // The manual's feedforward example: train from the CoNLL-U reader with
the default
+ // settings, save the model, load it back, and parse with a beam of four.
The parse
+ // of the reloaded model must equal the parse of the model in memory.
+ final Path treebank = dir.resolve("train.conllu");
+ Files.writeString(treebank, repeat(CONLLU, REPETITIONS),
StandardCharsets.UTF_8);
+ final Path modelFile = dir.resolve("en-depparse-ff.bin");
+ final InputStreamFactory trainingData = () ->
Files.newInputStream(treebank);
+
+ FeedforwardDependencyModel model;
+ try (ObjectStream<DependencySample> samples =
+ new ConlluDependencySampleStream(trainingData, ConlluTagset.U)) {
+ model = FeedforwardDependencyTrainer.train(
+ samples, FeedforwardDependencyTrainer.Settings.defaults());
+ }
+ try (OutputStream out = Files.newOutputStream(modelFile)) {
+ model.serialize(out);
+ }
+
+ DependencyParser reloaded = new FeedforwardDependencyParser(
+ FeedforwardDependencyModel.load(modelFile), BEAM_SIZE);
+ final DependencyTree expected =
+ new FeedforwardDependencyParser(model, BEAM_SIZE).parse(ALICE_TOKENS,
ALICE_TAGS);
+ final DependencyTree parse = reloaded.parse(ALICE_TOKENS, ALICE_TAGS);
+ assertEquals(expected, parse);
+ assertEquals(ALICE_TOKENS.length, parse.size());
+ assertEquals(ALICE_TREE.root(), parse.root());
+ }
+
+ /**
+ * Repeats CoNLL-U content so a small fixture gives the trainers enough
evidence.
+ *
+ * @param conllu The content to repeat; sentences are separated by a blank
line.
+ * @param times How often to repeat it.
+ * @return The repeated content. Never {@code null}.
+ */
+ private static String repeat(String conllu, int times) {
Review Comment:
Uses `String.repeat`, the local is `feedforwardModel`, and the test asserts
`ALICE_TREE` after the save and reload.
##########
opennlp-eval-tests/src/test/java/opennlp/tools/eval/UniversalDependencyParserEval.java:
##########
@@ -50,9 +53,12 @@
* and both again over the tokens that are not tagged {@code PUNCT}, the
customary
* reporting convention for Universal Dependencies.</p>
*
- * <p>The parser is trained with a feature cutoff of {@value #FEATURE_CUTOFF}
and
- * evaluated on English, German, Spanish, and French. The English cross
validation uses
- * {@value #ENGLISH_FOLDS} folds of the training split.</p>
+ * <p>The classical parser is trained with a feature cutoff of {@value
#FEATURE_CUTOFF}
Review Comment:
Now `{@link DependencyParserME}` and "the English EWT and Spanish AnCora
development splits".
##########
opennlp-docs/src/docbkx/dependency.xml:
##########
@@ -91,6 +93,32 @@ opennlp DependencyParserME en-depparse.bin <
tagged-sentences.txt]]>
</para>
</section>
+ <section xml:id="tools.dependency.feedforward">
+ <title>Feedforward parser</title>
+ <para>
+ The feedforward trainer has its own versioned model
format. This example trains,
+ saves, reloads, and creates a parser with a beam size
of four:
+ <programlisting language="java">
+<![CDATA[FeedforwardDependencyModel model;
+try (ObjectStream<DependencySample> samples =
+ new ConlluDependencySampleStream(trainingData, ConlluTagset.U)) {
+ model = FeedforwardDependencyTrainer.train(
+ samples, FeedforwardDependencyTrainer.Settings.defaults());
+}
+try (OutputStream out = Files.newOutputStream(Path.of("en-depparse-ff.bin"))) {
+ model.serialize(out);
+}
+
+DependencyParser parser = new FeedforwardDependencyParser(
+ FeedforwardDependencyModel.load(Path.of("en-depparse-ff.bin")), 4);]]>
+ </programlisting>
+ </para>
+ <para>
Review Comment:
The section is down to the example (now `serialize(Path)` and the `Path`
constructor, like the section above), one sentence on greedy versus beam and
thread sharing, and the evaluation paragraph names which parser is scored on
which split.
> Add a feedforward neural dependency parser
> ------------------------------------------
>
> Key: OPENNLP-1991
> URL: https://issues.apache.org/jira/browse/OPENNLP-1991
> Project: OpenNLP
> Issue Type: Improvement
> Reporter: Kristian Rickert
> Assignee: Kristian Rickert
> Priority: Major
> Time Spent: 10m
> Remaining Estimate: 0h
>
> Add a feedforward neural implementation of the dependency parser API
> introduced by OPENNLP-547. Use OpenNLP’s standard model container, artifact
> serializers, TrainingParameters, factory, and CLI conventions. Include
> deterministic training, greedy and beam decoding, model round-trip tests, and
> held-out Universal Dependencies evaluation. Address the deferred review
> feedback on shared scoring/training logic and parser-owned caches. Implement
> this without new dependencies.
> The list above may still get this ticket to further split. For now this is a
> placeholder to capture the code that was moved out of OPENNLP-547. I'll edit
> after 547 is merged.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
