Automatic annotation tools (part-of-speech taggers, lemmatizers, etc.) are often used when producing gold-standard annotations: it is often faster — and more convenient — to review and correct annotations created by a software tool than it is to annotate everything from scratch. However, for non-standard language data (e.g. historical texts, less-resourced languages), suitably trained annotation software often doesn’t exist, so it might be desirable to (re-)train an annotation program with new manually annotated data as it becomes available.

For this reason, CorA offers the functionality to embed automatic annotation tools and call them directly from the GUI.

---

To manage external annotators, change to the tab “Administration” while logged in to CorA with an administrator account.

• Create a new annotator by clicking the “Add Tagger” button at the top of the page. You will be prompted for the tagger’s name first, and can edit individual settings afterwards.

You can manage existing annotators in the “Automatic Annotators” section of the admin page. Individual settings can be changed by clicking on “Options…”, namely:

• PHP class: Embedding an external annotation tool requires a PHP class that acts as an interface between CorA and the external tool. The name of the class goes in this field. See below for details on tagger interfaces.

• Options: Enter an arbitrary list of key/value pairs as options for the PHP annotator class. Which options are available depends entirely on the PHP class that you use. Click on or to add or remove rows from the option list, respectively.

• Individually trained: Check this box if the tagger should be (re-)trained on documents within CorA. This requires the PHP class to support this, and the configuration option “external_param_dir” to be set to an existing directory with read/write permissions for the web server process. Leave this unchecked if the tagger already is or does not need to be trained.

• Tagset links: Specify which tagset(s) are required to be present for this tagger to work. The tagger will be available for all documents that have at least the tagsets specified here associated with them.

By default, annotators will operate on the “ascii” form of the modernized tokens (cf. CorA’s document model).

Tagger interfaces

As said above, embedding an external annotation tool requires a PHP class that acts as an interface between CorA and the external tool. Currently, the following interfaces are already part of the CorA distribution, i.e. you can use them “out-of-the-box”:

Lemmatizer

Acts as a simple wrapper around a lemmatization script. Supports the following options:

• bin: path to the lemmatization script (required); see bin/cora_lemmatize.perl for an example script that can be used with this class
• filter_unknown (boolean – 0 or 1): if 1, filters annotations where the lemmatization script returned ““; defaults to 1
• lowercase_all (boolean – 0 or 1): if 1, all input is lowercased; defaults to 0
• par: path to the parameter file (required if the lemmatizer should not be individually trained)
• use_norm (boolean – 0 or 1): if 1, use the normalization as input; defaults to 0
• use_pos (boolean – 0 or 1): if 1, make use of POS annotation when lemmatizing; defaults to 1

RFTagger

Acts as a wrapper around RFTagger, a part-of-speech tagger developed by Helmut Schmid and Florian Laws. Supports the following options:

• annotate: path to the rft-annotate binary (required)
• flags: additional flags for RFTagger; defaults to -c 2 -q
• lowercase_all (boolean – 0 or 1): if 1, all input is lowercased; defaults to 0
• minimum_span_size: when re-training on a document that has “gaps” in the annotation, only chunks of continuously annotated tokens of at least this size are used in training; defaults to 5
• par: path to the parameter file (required if the tagger should not be individually trained)
• train: path to the rft-train binary (required if the tagger should be individually trained)
• use_layer: whether to use ‘ascii’, ‘trans’, or ‘utf’ forms of token; can also be set to ‘norm’ to use the normalization as input; defaults to ‘ascii’
• wc: path to RFTagger wordclass file (required)

DualRFTagger

Combines two instances of RFTagger: one using a provided parameter file; the other one individually trained. During annotation, both instances are called. The tag returned by the individually trained model is chosen if either (a) the input token was already seen during training (with a number of occurences greater than a certain threshold) and its tag is not “?”; or (b) the fixed model returned “?” as its tag. Otherwise, the tag from the fixed model is chosen.

Supports the same options as RFTagger, and additionally:

• threshold: the threshold for “trusting” the individually trained model; defaults to 1

Writing your own interface

If you’d like to integrate some external tool that is not covered by one of the interface classes above, you need to write your own PHP class for it.

All annotation interfaces reside in <cora-dir>/lib/annotation/. There is a base class “AutomaticAnnotator” (in AutomaticAnnotator.php) from which all other interfaces should inherit. Place your custom interface, say, “FooTagger”, in a file FooTagger.php in this directory. You can then create a new annotator in CorA with a PHP class of “FooTagger” which uses this interface.

Please refer to the code documentation for more details on what your custom class should implement.