Hello,
We have just upgraded from Freeling 3.1 to 4.0. We want to use the semantic graph, so we have changed de dependecy Anaylyzer used, from Txala to a mixed depTxala-Treeler analysis (as the example java file shows).
Is this change mandatory to use semantic graph and dependency analysis?
What do we improve with the annexed change?
// depTxala.analyze( doc ); //3.1
depTxala.completeParseTree(doc); //4.0
depTreeler.analyze(doc); //4.0
Now, the labels I get are a mix of the depTxala labels (.\FreeLing-4.0\doc\grammars\ca+esLABELINGtags) and the Ancora labels (spanish, Ancora labels 2.0 or 1.0) ?
We have noticed that some dependencies have change, from 3.1. As an example, in the sentence 'Dame mis recibos de teléfono del último trimestre', 'trimestre' used to depend on 'recibos' (as 'teléfono' did), but now it depends on 'teléfono'. Syntactically, this dependency is wrong.
Is this change due to the dependency analyzer change?
What can we do to help Freeling to analyze the dependences right?
Semantic graph building
Semantic graph building requires SRL and coreferences.
For Spanish, SRL is provided by depTreeler, and Coreferencies require depTxala parse trees (but not dependency trees).
Thus, you need to run completeParseTree build the constituency tree that coreferences will use.
Also, you need to run depTreeler to get dependencies and SRL
But:
- The constituency tree should be 100% syntactic (so, labels like grup-np, etc) and have no dependency relations (subj, dobj, etc). You can ignore the constituency tree, since it is build only to meet corefencence needs.
- The dependency tree and SRL should be 100% ancora labels, since depTreeler is trained on that corpus.
Dependencies are different because in 3.1, depTxala was used, and in 4.0, depTreeler is used.
If you want only the tree, you can use either one or the other, but if you want to use semantic graph, there is no choice, since it needs SRL (only available in depTreeler).
Alternatively, you can also generate semantic graphs running only depTreeler (that is , without running neither depTxala nor coreferences), but they will have different entity nodes for different mentions of the same entity.