Why does the property settings for the StanfordCoreNlp pipeline matters when building dependency parse tree? - stanford-nlp

Question

I have been playing around with Stanford-CoreNLP and I figured out that building a dependency parse tree with the following code
String text = "Are depparse and parse equivalent properties for building dependency parse tree?"
Properties props = new Properties();
props.setProperty("annotators", "tokenize, ssplit, parse, lemma, ner");
StanfordCoreNLP pipeline = new StanfordCoreNLP(props);
Annotation document = new Annotation(text);
pipeline.annotate(document);
List<CoreMap> sentences = document.get(CoreAnnotations.SentencesAnnotation.class);
for (CoreMap sentence : sentences) {
SemanticGraph graph = sentence.get(SemanticGraphCoreAnnotations.BasicDependenciesAnnotation.class);
System.out.println(graph.toString(SemanticGraph.OutputFormat.LIST ));
}
is outputting
root(ROOT-0, tree-11)
cop(tree-11, Are-1)
amod(properties-6, depparse-2)
cc(depparse-2, and-3)
conj(depparse-2, parse-4)
compound(properties-6, equivalent-5)
nsubj(tree-11, properties-6)
case(dependency-9, for-7)
compound(dependency-9, building-8)
nmod(properties-6, dependency-9)
amod(tree-11, parse-10)
punct(tree-11, ?-12)
However this code
String text = "Are depparse and parse equivalent properties for building dependency parse tree?"
Properties props = new Properties();
props.setProperty("annotators", "tokenize, ssplit, lemma, ner, depparse");
StanfordCoreNLP pipeline = new StanfordCoreNLP(props);
Annotation document = new Annotation(text);
pipeline.annotate(document);
List<CoreMap> sentences = document.get(CoreAnnotations.SentencesAnnotation.class);
for (CoreMap sentence : sentences) {
SemanticGraph graph = sentence.get(SemanticGraphCoreAnnotations.BasicDependenciesAnnotation.class);
System.out.println(graph.toString(SemanticGraph.OutputFormat.LIST ));
}
outputs
root(ROOT-0, properties-6)
cop(properties-6, Are-1)
compound(properties-6, depparse-2)
cc(depparse-2, and-3)
conj(depparse-2, parse-4)
amod(properties-6, equivalent-5)
case(tree-11, for-7)
amod(tree-11, building-8)
compound(tree-11, dependency-9)
amod(tree-11, parse-10)
nmod(properties-6, tree-11)
punct(properties-6, ?-12)
So why am I not getting the same outputs with those two methods? Is it possible to change the later code to be equivalent to the first code because loading the constituency parser as well makes the parsing so slow? And how would you recommend setting the properties to get the most accurate dependency parse tree?

Answer 1

The constituency parser (parse annotator) and dependency parser (depparse annotator) are actually completely different models and code paths. In one case, we're predicting a constituency tree and converting it to a dependency graph. In the other case, we are running a dependency parser directly. In general, depparse is expected to be faster (O(n) vs O(n^3)) and more accurate at producing dependency trees, but will not produce constituency trees.