CheetahTop

Cheetah is a grammar for German, meant to be used with either the LKB or PET. It is comprised of a hand-written core grammar and a lexicon derived from the Tiger treebank, and a 25k sentence [incr tsdb()] treebank, converted from the same Tiger treebank.

The core grammar

The core grammar is heavily inspired by the following papers:

• Hinrichs, E. and Nakazawa, T. 1994. Linearizing AUXs in German verbal complexes. In: German in Head-Driven Phrase Structure Grammar.

• Müller, S. 2002. Complex predicates: Verbal complexes, resultative constructions, and particle verbs in German.

• Crysmann, B. 2003. On the efficient implementation of German verb placement in HPSG. In Proceedings of RANLP 2003, pages 112–116.

• Crysmann, B. 2005. Relative Clause Extraposition in German: An Efficient and Portable Implementation. Research on Language & Computation, 3(1):61–82.

Currently, the following phenomena are supported:

• Mittelfeld scrambling, verbal clusters
• Extraposition of complements, adjuncts, relative phrases and comparatives
• Certain forms of ellipsis
• Passives

The core grammar contains a core lexicon, roughly containing all words from closed word classes. It contains 550 lemmas.

More information on the internals of the core grammar can be obtained if you contact me directly. Some documentation can also be found in my PhD thesis (in preparation).

Deep lexical acquisition

To obtain coverage, a deep lexical acquisition step has been carried out on sentence 1-45000 of the Tiger treebank, mapping lemmas to detailed lexical types. For verbs, around 200 distinct types are found for verbs. Some rough stats on this step are:

Part-of-speech	Lemmas	Lexical entries	Inflection triples
Verbs	4543	9235	18745
Nouns	33835	34821	51303
Names	12445	12783	na
Adjectives	7318	8018	50480
Adverbs	2654	4577	na

The Savannah treebank

The grammar does not output proper MRSs (although it uses the same format), but Tiger-like dependencies. By re-parsing the raw text from the Tiger treebank, and annotating the best reading, an [incr tsdb()] treebank has been created. Of the first 45k sentences in Tiger, 25k receive a dependency f-score higher than 0.9.

Due to licensing issues, you can obtain the Savannah treebank only if you or your institute has a license for the original Tiger treebank. However, getting it is trivial for non-commercial use (just click the OK button): License agreement.

Results

Coverage of unseen Tiger text is about 60%, of which around 35% is exact match (top-500) or 17% (top-1).

Technical details

A typical run of the grammar would look as follows:

 cat input.yy | cheap -cm -yy -default-les=traditional -mrs=simple -packing=15 -nsolutions=1 -sm=sav-45k.mem -cp=300 german.grm 

The YY file will restrict with respect to the part of speech indicated in the YY file, which is mandatory. Refer to types_lexical.tdl to see which parts of speech ought to be used. An example of an input YY file is:

(0, 0, 1, <0:3>, 1, "nach", 0, "null", "ADP" 1.0) (1, 1, 2, <5:20>, 1, "meinungsumfragen", 0, "null", "NN" 1.0) (2, 2, 3, <22:26>, 1, "liegt", 0, "null", "VFIN" 1.0) (3, 3, 4, <28:30>, 1, "der", 0, "null", "DET" 1.0) (4, 4, 5, <32:41>, 1, "parteilose", 0, "null", "ADJ" 1.0) (5, 5, 6, <43:54>, 1, "self-mademan", 0, "null", "NE" 1.0) (6, 6, 7, <56:58>, 1, "gut", 0, "null", "ADV" 1.0) (7, 7, 8, <60:61>, 1, "im", 0, "null", "ADP" 1.0) (8, 8, 9, <63:68>, 1, "rennen", 0, "null", "NN" 1.0) (9, 9, 10, <70:71>, 1, "um", 0, "null", "ADP" 1.0) (10, 10, 11, <73:75>, 1, "den", 0, "null", "DET" 1.0) (11, 11, 12, <77:86>, 1, "chefsessel", 0, "null", "NN" 1.0) (12, 12, 13, <88:89>, 1, "im", 0, "null", "ADP" 1.0) (13, 13, 14, <91:96>, 1, "weißen", 0, "null", "NE" 1.0) (14, 14, 15, <98:101>, 1, "haus", 0, "null", "NE" 1.0) (15, 15, 16, <103:105>, 1, "mit", 0, "null", "ADP" 1.0) (16, 16, 17, <107:109>, 1, "dem", 0, "null", "DET" 1.0) (17, 17, 18, <111:121>, 1, "amtierenden", 0, "null", "ADJ" 1.0) (18, 18, 19, <123:133>, 1, "präsidenten", 0, "null", "NN" 1.0) (19, 19, 20, <135:140>, 1, "george", 0, "null", "NE" 1.0) (20, 20, 21, <142:145>, 1, "bush", 0, "null", "NE" 1.0) (21, 21, 22, <147:149>, 1, "und", 0, "null", "KON" 1.0) (22, 22, 23, <151:153>, 1, "dem", 0, "null", "DET" 1.0) (23, 23, 24, <155:164>, 1, "demokraten", 0, "null", "NN" 1.0) (24, 24, 25, <166:169>, 1, "bill", 0, "null", "NE" 1.0) (25, 25, 26, <171:177>, 1, "clinton", 0, "null", "NE" 1.0)

Other remarks

• There is the possibility to generate in principle, but this functionality has never been tested on a large scale.

Publications

• Bart Cramer and Yi Zhang. 2009. Constructon of a German HPSG grammar from a detailed treebank. In: Proceedings of the ACL 2009 Grammar Engineering across Frameworks workshop, pages 37-45, Singapore, Singapore.
• Bart Cramer and Yi Zhang. 2010. Constraining robust constructions for broad-coverage parsing with precision grammars. In: Proceedings of COLING-2010.

@inproceedings{CramerZhang09,
  title={{Construction of a German HPSG grammar from a detailed treebank}},
  author={Cramer, B. and Zhang, Y.},
  booktitle={Proceedings of the GEAF workshop ACL-IJCNLP 2009},
  year={2009},
  pages={37--45}
}

@inproceedings{CramerZhang10,
  title={{Constraining robust constructions for broad-coverage parsing with precision grammars}},
  author={Cramer, B. and Zhang, Y.},
  booktitle={Proceedings of COLING 2010},
  year={2010},
  pages={{To appear}}
}