I'm trying to use simple rules/patterns defined over a dependency graph to extract very basic informations from sentences (e.g., triples such as subject->predicate->object). I started using StanfordNLP since it was easy to set up and utlizes the GPU for better performance. However, I've noticed that for some sentences, the resulting dependency graph looked not as I would have expected -- I'm no expert though. I therefore tried two other solutions: spaCy and Stanford CoreNLP (I understand that these are maintained by different groups?)
For the example sentence "Tom made Sam believe that Alice has cancer." I've printed the dependencies for all three approaches. CoreNLP and spaCy yield the same dependencies, and they are different from the ones of StanfordNLP. Hence, I'm inclined to swich to CoreNLP and spaCy (another advantage would be that they come with NER out of the box).
Does anyone have some more experience or feedback that would help where to go from here? I don't expect that CoreNLP and spaCy will always yield in the same dependency graphs, but in the example sentence, considering Sam as obj as StandfordNLP is doing compared to being nsubj (CoreNLP, spaCy) seems to be a significant difference
Format:
token dependency_tag parent_token
StanfordNLP
Tom nsubj made
made ROOT ROOT
Sam obj made
believe ccomp made
that mark has
Alice nsubj has
has ccomp believe
cancer obj has
. punct made
CoreNLP
Tom nsubj made
made ROOT ROOT
Sam nsubj believe
believe ccomp made
that mark has
Alice nsubj has
has ccomp believe
cancer dobj has
. punct made
spaCy
Tom nsubj made
made ROOT ROOT
Sam nsubj believe
believe ccomp made
that mark has
Alice nsubj has
has ccomp believe
cancer dobj has
. punct made
Not sure how to address your questions but I'd recommend you carefully read the documentation for the Stanford CoreNLP: https://nlp.stanford.edu/software/lex-parser.shtml
Within the package there are several grammatical and dependency parsers that you can use. Just looking at the grammatical parses, there is an option to retrieve k-best parses and if you process dependencies on them you will most likely get different dependencies for each.
This has to do both with inaccuracies in the parser and ambiguities in natural language.