Citar - A simple Trigram HMM part-of-speech tagger

== Introduction ==

Citar is a simple part-of-speech tagger, based on a trigram Hidden
Markov Model (HMM). It (partly) implements the ideas set forth in
[1]. Citar is written in C++.

Citar is licensed under the GNU Lesser General Public License
version 3.0.

== Warning ==

The Citar API will be highly unstable for the first few versions!

== Building Citar ==

Builing Citar requires a C++ standard library with TR1 extensions,
such as a recent version of libstdc++ as included with GNU g++. This
release was tested with g++ 4.3.2. cmake is used for creating build
infrastructure.

You can create the build infrastructure by running "ccmake ." in the
top-level Citar directory. This will allow you to configure various
settings. The WITH_TRIGRAM_CACHE setting is used to enable/disable the
trigram cache for linear interpolation smoothing. This may give a
performance gain in some situations, but is currently not thread-safe.

After configuring Citar with cmake you can invoke "make" on Unix
systems to build Citar. Command-line utilities for training and
evaluating the tagger will be produced. Compilation will also produce
the 'libsitar.a' library, which you can use to integrate the tagger in
your own programs.

== Training ==

The language model and lexicon can be created with the 'train' utility:

---
$ ./train corpus-train lexicon ngrams
---

This will create the 'lexicon' and 'ngrams' files. The trainer will read
corpora in the Brown format (one sentence per line, words and tags are
separated with a forward slash). You can now test the tagger with the
command-line 'tag' utility, which reads tokenized sentences from the
standard input and prints the most probable tag sequence:

---
$ echo "The cat is on the mat ." | ./tag lexicon ngrams
The/AT cat/NN is/BEZ on/IN the/AT mat/NN ./.
---

== Authors ==

Daniel de Kok <[email protected]>

== FAQ ==

- "What's up with the name?"

  Citar, it is not an abbreviation. If you do prefer abbreviations,
  let's make it "C++ sImple TAgging Redux" :).

[1] TnT - a statistical part-of-speech tagger, Thorsten Brants, 2000