Version 3 of the toolkit introduces significant changes to the package, largely in response to improvements in the underlying NLP annotators.
we now use the official Python version of CoreNLP. It supports a much larger number of languages, but at the moment only allows for dependency parsing (no NER or coreferences). This removes the need for a Java backend and significantly simplifies the package.
we now have a Python packaged, cleannlp, that requires spacy and corenlp. This removes the need to include python scripts with the R package and should make it easier to install the toolkit.
all functions work with data in memory. Data are never stored on disk by the package. This makes it much easier to work with large collections of small documents.
the tokens and dependency tables are returned pre-combined. While there is some academic justification for separating them, because they always have the same number of rows, there is no practical reason to do so.
the annotation object is now returned as a plain, unclassed list. The get functions are no longer included or needed (they added a lot of complexity for little benefit).
removed all internal usage of dplyr. Output data frames still include a “tibble” class indicator, so if users import dplyr the pretty printing will be preserved.
removed phantom empty columns; only columns with data are returned.
removed many options that were no longer relevant to the underlying algorithms.
Users with scripts based on the previous version of cleanNLP will need to modify them to match the new semantics. We believe the small changes required will make the toolkit easier to both install and use.
This is a major re-structuring of the cleanNLP package. The primary changes include:
the new udpipe backend, which gives tokenization, POS-tags, lemmatization,and dependency parsing with no external dependencies
all functions return results in memory; to store data on disk, users need to save the output manually
use a character vector named ‘id’ as the document id (it was previously an integer index); this is to conform to the text interchange format (tif)
functions now use the prefix ‘cnlp_’, following the convention of packages such as stringi
the cnlp_get_tfidf function now returns a named sparse matrix in lieu of a named list
There are also many internal changes, primarily to deal with the new spaCy (2.0) version and to make the use of udpipe more naturally.
In this version, the internal mechanisms for running the tokenizers backend have been changed. We are now directly calling the stringi functions with options that better mimic those of the the spaCy and CoreNLP backends. Despite the lack of dependency on the tokenizers package, we will continue to use the name “tokenizers” for the backend to maintain backwards consistency.
As part of the change to custom stringi function, we now also support setting the locale as part of initalizing the tokenizers backend. This allows for an easy way of tokenizing text where custom spaCy or coreNLP models do not yet exist.
There is currently a pre-release version of spaCy version 2.0.0. The current version has been tested and runs smoothly with cleanNLP. The new neural network models are sufficently faster and more accurate; we suggest migrating to the version 2 series as it becomes stable for production.
This version contains many internal changes to the way that external libraries are called and referenced in order to comply with goodpractice::gp(). Two important user-facing changes include:
The function annotate has been changed to
run_annotations in order to avoid a conflict with
ggplot2::annotate
The function get_token has new options for producing
a single joined table with dependencies and entities. This should make
it easier to work with the output for users needed more than lemmas and
POS-tags but not requiring deeper table joins.
This update contains several major changes, include:
document and sentence ids now start at 1
download function checks and warns if Java files are already downloaded
table joins inside of get_document() no longer produce verbose output
get_token() now has an option, FALSE by default, for whether sentence ROOTS should be returned
the speed parameter to init_coreNLP() has been renamed as anno_level