This function must be run before annotating text with the corenlp backend. It sets the properties for the corenlp engine and loads the file using rJava interface provided by reticulate. See Details for more information about the anno_level codes.
cnlp_init_corenlp(language, anno_level = 2, lib_location = NULL, mem = "6g", verbose = FALSE)
a character vector describing the desired language; should be one of: "ar", "de", "en", "es", "fr", or "zh".
integer code. Sets which annotators should be loaded, based on on how long they take to load and run. anno_level 0 is the fastest, and anno_level 8 is the slowest. See Details for a full description of the levels
a string giving the location of the corenlp java files. This should point to a directory which contains, for example the file "stanford-corenlp-*.jar", where "*" is the version number. If missing, the function will try to find the library in the environment variable corenlp_HOME, and otherwise will fail. (Java model only)
a string giving the amount of memory to be assigned
to the rJava engine. For example, "6g" assigned 6
gigabytes of memory. At least 2 gigabytes are
recommended at a minimum for running the corenlp
package. On a 32bit machine, where this is not
possible, setting "1800m" may also work. This
option will only have an effect the first
init_backend is called for the corenlp
backend, and also will not have an effect if the
java engine is already started by another process.
boolean. Should messages from the pipeline be written to the console or suppressed?
Currently available anno_level codes are integers from 0 to 8. Setting anno_level above 2 has no additional effect on the German and Spanish models. Setting above 1 has no effect on the French model. The available anno_level codes are:
"0" runs just the tokenizer, sentence splitter, and part of speech tagger. Extremely fast.
"1" includes the dependency parsers and, for English, the sentiment tagger. Often 20-30x slower than anno_level 0.
"2" adds the named entity annotator to the parser and sentiment tagger (when available). For English models, it also includes the mentions and natlog annotators. Usually no more than twice as slow as anno_level 1.
"3" add the coreference resolution annotator to the anno_level 2 annotators. Depending on the corpus, this takes about 2-4x longer than the anno_level 2 annotators
We suggest starting at anno_level 2 and down grading to 0 if your corpus is particularly large, or upgrading to 3 if you sacrifice the slowdown. If your text is not formal written text (i.e., tweets or text messages), the anno_level 0 annotator should still work well but anything beyond that may be difficult. Semi-formal text such as e-mails or transcribed speech are generally okay to run for all of the levels.