This distribution contains the following directories and files:

• doc
  • This directory contains documentation of the program.
• src
  • This directory contains source code and a Makefile.
  • You also need to clone source code from the following repositories to lib/parsesgml, lib/letterfunc and lib/hashmap respectively:
    • https://github.com/kuhumcst/parsesgml
    • https://github.com/kuhumcst/letterfunc
    • https://github.com/kuhumcst/hashmap
  • NOTE: Fetching libraries and compiling the code can be done automatically by running makecstlemma.bash.
  • ALTERNATIVELY: Place the three included source directories in the same root directory that the cstlemma directory is in and then run make inside cstlemma/src.
• Changelog
  • A document describing changes between versions.
• COPYING
  • The full text of the GNU public licence.
• README.md
  • This file.

CSTLEMMA has been compiled and run on the following platforms:

Both 32 and 64 bit versions can be made.

Installation

• Linux/Mac:
  1. Download (e.g. git clone) cstlemma, parsesgml, letterfunc and hashmap. If you are going to use the Makefile that comes with cstlemma, locate each of these packages in separate subdirectories under the same directory, and call these subdirectories cstlemma, parsesgml, letterfunc and hashmap, respectively. You can run makecstlemma.bash to do all of this automatically.
  2. Change directory to the 'cstlemma/src' directory.
  3. Run 'make' (both Linux and Mac) or 'make all' (only Linux). To get rid of object files, run
  4. 'make clean'.

Running

For running the CST lemmatiser you need as a minimum a file containing flex rules. The absolute minimal set of flex rules is the empty set, in which case the lemmatiser assumes that all words in your input text are perfectly lemmatised already.

Thus, for checking that the lemmatiser runs OK, you could do the following:

    touch my_empty_rule_file
    ./cstlemma -L -f my_empty_rule_file -i my_text_file.txt

This would create a file my_text_file.txt.lemma that has two tab-separated columns: the left column contains a word from your text and the right column contains the same word, converted to lower-case. The -L option tells the program lemmatise (as opposed to generating flex rules or creating a binary dictionary). The -f and -i options tell the program which rules and which input text to read.

The lemmatiser "cstlemma" can generate flex rules from a full-form dictionary, but a better way to obtain flex rules for use by cstlemma is to use the program called "affixtrain" (https://github.com/kuhumcst/affixtrain). The flex rules that cstlemma can produce only look at words from the end ("suffix"-oriented). Affixtrain, on the other hand, can look at the beginning, the end and in several places inside a word, all at once. For languages like German and Dutch, were morphological changes deep inside words often follow strict rules, affixtrain has a clear advantage over the old suffix based algorithm implemented in cstlemma, but also languages that only have suffix morphology can have words that carry important information in other places than near the end about word class and therefore morphology. A disadavantage of training flex rules with affixtrain is that it can take a long time, perhaps days if the training data consists of millions of unique full form - lemma pairs.

The full-form dictionary used to train flex rules can also be used to generate a binary dictionary, which the program can use to even better lemmatise your input text. For this task you can use cstlemma (with the -D option).

Online availability

CSTLEMMA is demonstrated at CST's website (http://ada.sc.ku.dk/tools/index.php?lang=en) and an integrated webservice in the CLARIN-DK infrastructure (https://clarin.dk/).

References

Bart Jongejan and Hercules Dalianis. 2009. Automatic training of lemmatization rules that handle morphological changes in pre-, in- and suffixes alike. In Proceedings of the Joint Conference of the 47th Annual Meeting of the ACL and the 4th International Joint Conference on Natural Language Processing of the AFNLP, pages 145–153. Association for Computational Linguistics. https://www.aclweb.org/anthology/P09-1017.pdf

Appraisal of lemmatization (author: ChatGPT)

There once was a word, quite a beast

Its endings were many, at least

But with lemmatization

It found liberation

And made perfect sense in the least

Contact info

For questions and remarks about the program, please feel free to contact us.

Our postal address is:

Center for Sprogteknologi
University of Copenhagen
Njalsgade 140
2300 Copenhagen S.
Denmark

On the internet, you can visit us at www.cst.ku.dk Here you can also try the CST lemmatiser for Danish and many other languages.