Institute of Formal and Applied Linguistics Wiki

Eesti keele puudepank (Google translate) (EKP)

• Downloadable on-line, part of Arborest project (puudepank)
• 8.12.2010 arborest.xml downloadable from the same site (same size, improved markup)
• http://vvv.cs.ut.ee/~kaili/Korpus/pindmine/

The EKP is freely downloadable from here in VISL or TIGER-XML format. Licensing terms are unknown.

EKP was created / coordinated (?) by Kaili Müürisep, Institute of Computer Science (Arvutiteaduse instituut), University of Tartu (Tartu Ülikool), Liivi 2, 50409 Tartu, Estonia.

• Website
  • http://vvv.cs.ut.ee/~kaili/Korpus/puud/ (Google translate)
• Data
  • no separate citation
• Principal publications
  • Kaili Müürisep, Tiina Puolakainen, Kadri Muischnek, Mare Koit, Tiit Roosmaa, Heli Uibo: A New Language for Constraint Grammar: Estonian. In: International Conference Recent Advances in Natural Language Processing. Proceedings, pp. 304-310, Borovets, Bulgaria, 2003.
• Documentation
  • File formats
  • The header of the TIGER-XML version of the treebank contains lists of various sorts of tags with brief explanation.

Mixed:

• 388 tailored sentences with movement verbs
• 732 sentences with movement verbs from the Estonian FrameNet corpus
• 175 sentences from the Arborest corpus
• 20 sentences of spoken language

All four parts of the treebank together contain 9491 tokens in 1315 sentences, yielding 7.22 tokens per sentence on average. No official training-test data split is defined. Due to the small size of the treebank and extraordinary domain diversity, a good test set should sample from all four parts of the treebank. This is the case of our HamleDT experimental data split, shown in the last two rows of the table.

The treebank is part of the Arborest project and VISL (Visual Interactive Syntax Learning). As such, it is based on Constraint Grammar (Fred Karlsson et al., 1995: Constraint Grammar – A Language-Independent System for Parsing Unrestricted Text. Mouton de Gruyter). All four parts are available in the TIGER-XML format. Two of them are also available in the VISL format.

The annotation contains lemmas, part of speech tags, morphosyntactic features, nonterminal labels and phrase structure. It is not clear whether (and to what degree) the annotation was performed or checked manually.

Note that the TIGER-XML format, despite being phrase-based, stores word order separately from structure and thus allows for nonprojectivities.

The first sentence of the corpus in the TIGER-XML format:

<s id="ratsep-13" ref="ratsep-1" source="id=ratsep-1" forest="1/1" text="Peeter aerutas üle väina saarele puhkama">
	<graph root="ratsep-13_501">
		<terminals>
			<t id="ratsep-13_1" word="Peeter" lemma="Peeter+0" pos="prop" morph="prop,sg,nom,.cap"/>
			<t id="ratsep-13_2" word="aerutas" lemma="aeruta+s" pos="v-fin" morph="main,indic,impf,ps3,sg,ps,af,.FinV"/>
			<t id="ratsep-13_3" word="üle" lemma="üle+0" pos="prp" morph="pre,.gen"/>
			<t id="ratsep-13_4" word="väina" lemma="väin+0" pos="n" morph="com,sg,gen"/>
			<t id="ratsep-13_5" word="saarele" lemma="saar+le" pos="n" morph="com,sg,all"/>
			<t id="ratsep-13_6" word="puhkama" lemma="puhka+ma" pos="v-inf" morph="main,sup,ps,ill,.Part"/>
			<t id="ratsep-13_7" word="." lemma="." pos="punc" morph="Fst"/>
		</terminals>
 
		<nonterminals>
			<nt id="ratsep-13_501" cat="VROOT">
				<edge label="STA" idref="ratsep-13_502"/>
			</nt>
			<nt id="ratsep-13_502" cat="fcl">
				<edge label="S" idref="ratsep-13_1"/>
				<edge label="P" idref="ratsep-13_2"/>
				<edge label="A" idref="ratsep-13_503"/>
				<edge label="A" idref="ratsep-13_5"/>
				<edge label="A" idref="ratsep-13_6"/>
				<edge label="FST" idref="ratsep-13_7"/>
			</nt>
			<nt id="ratsep-13_503" cat="pp">
				<edge label="H" idref="ratsep-13_3"/>
				<edge label="D" idref="ratsep-13_4"/>
			</nt>
		</nonterminals>
	</graph>
</s>

Nonprojectivities in EKP are very rare. Only 7 out of the 9491 tokens are attached nonprojectively (0.074%).

There is a constraint grammar parser for Estonian by Kaili Müürisep. I am not aware of any published evaluation of parsing accuracy. However, I am not sure that the treebank described here is not just output of the parser.