Prague Arabic Dependency Treebank (PADT)

• Original PADT 1.0 as distributed by the LDC
• CoNLL 2006
• CoNLL 2007

The CoNLL 2007 version reportedly improves over CoNLL 2006 in quality of morphological annotation. Both CoNLL versions miss important parts of the original PADT annotation (see below).

The original PADT 1.0 is distributed by the LDC under the catalogue number LDC2004T23. It is free for LDC members 2004, price for non-members is unknown (contact LDC). The license in short:

• non-commercial research usage
• no redistribution
• cite one publication in publications; also include: “The PADT 1.0 has been developed by the Institute of Formal and Applied Linguistics and the Center for Computational Linguistics (see http://ufal.mff.cuni.cz/).”

The CoNLL 2006 and 2007 versions are obtainable upon request under similar license terms. Their publication in the LDC together with the other CoNLL treebanks is being prepared.

PADT was created by members of the Institute of Formal and Applied Linguistics (Ústav formální a aplikované lingvistiky, ÚFAL), Faculty of Mathematics and Physics (Matematicko-fyzikální fakulta), Charles University in Prague (Univerzita Karlova v Praze), Malostranské náměstí 25, Praha, CZ-11800, Czechia. The CoNLL conversion of the treebank was prepared by Otakar Smrž, one of the key authors of the original PADT.

Newswire text from press agencies (Agence France Presse, Ummah, Al Hayat, An Nahar, Xinhua 2001-2003).

According to their website, the original PADT 1.0 contains 113,500 tokens annotated analytically. The CoNLL 2006 version contains 59752 tokens in 1606 sentences, yielding 37.21 tokens per sentence on average (CoNLL 2006 data split: 54379 tokens / 1460 sentences training, 5373 tokens / 146 sentences test). The CoNLL 2007 version contains 116,793 tokens in 3043 sentences, yielding 38.38 tokens per sentence on average (CoNLL 2007 data split: 111,669 tokens / 2912 sentences training, 5124 tokens / 131 sentences test).

As noted in (Nivre et al., 2007), “the parsing units in this treebank are in many cases larger than conventional sentences, which partly explains the high average number of tokens per sentence.”

• Website
  • http://ufal.mff.cuni.cz/padt/
• Data
  • Jan Hajič, Otakar Smrž, Petr Zemánek, Petr Pajas, Jan Šnaidauf, Emanuel Beška, Jakub Kráčmar, Kamila Hassanová: Prague Arabic Dependency Treebank 1.0 (LDC2004T23). Linguistic Data Consortium, Philadelphia, USA, 2004. ISBN 1-58563-319-4.
• Principal publications
  • Jan Hajič, Otakar Smrž, Petr Zemánek, Jan Šnaidauf, Emanuel Beška: Prague Arabic Dependency Treebank: Development in Data and Tools. In: Proceedings of the NEMLAR International Conference on Arabic Language Resources and Tools, pages 110-117, Cairo, Egypt, 2004.

The original PADT 1.0 is distributed in the FS format. The CoNLL versions are distributed in the CoNLL-X format. The original PADT contains more information than the CoNLL version. There is morphological annotation (tags and lemmas) both manual and by a tagger (while only manual is in CoNLL data), glosses etc. However, the most important piece of information that got lost during the conversion to CoNLL is the FS attribute called parallel. It distinguishes conjuncts from shared modifiers of coordination and thus the syntactic structure is incomplete without it.

Word forms and lemmas are vocalized, i.e. they contain diacritics for short vowels as well as consonant gemination and a few other things. The CoNLL 2006 version includes Buckwalter transliteration of the Arabic script (in the same column as Arabic, attached to the Arabic form/lemma with an underscore character).

Note that tokenization of Arabic typically includes splitting original words (inserting spaces between letters), not just separating punctuation from words. Example: وبالفالوجة = wabiālfālūjah = wa/CONJ + bi/PREP + AlfAlwjp/NOUN_PROP = and in al-Falujah. In PADT, conjunctions and prepositions are separate tokens and nodes.

The original PADT 1.0 uses 10-character positional morphological tags whose documentation is hard to find. The CoNLL 2006 version converts the tags to the three CoNLL columns, CPOS, POS and FEAT, most of the information being encoded as pipe-separated attribute-value assignments in FEAT. There should be a 1-1 mapping between the PADT positional tags and the CoNLL 2006 annotation. The CoNLL 2007 version uses a tag conversion different from CoNLL 2006. Both CoNLL distributions contain a README file with a brief description of the parts of speech and features. Use DZ Interset to inspect the two CoNLL tagsets. Also look at the Elixir FM online interface for a later development of the morphological analyzer created along with PADT.

The guidelines for syntactic annotation are documented in the PADT annotation manual (only peculiarities of Arabic are documented, otherwise it is referenced to the annotation manual for the Czech treebank). The list and brief description of syntactic tags (dependency relation labels) can be found in (Smrž, Šnaidauf and Zemánek, 2002).

The first two sentences of the CoNLL 2006 training data:

The first sentence of the CoNLL 2006 test data:

The first sentence of the CoNLL 2007 training data:

The first sentence of the CoNLL 2007 test data:

Nonprojectivities in PADT are rare. Only 431 of the 116,793 tokens in the CoNLL 2007 version are attached nonprojectively (0.37%).

The results of the CoNLL 2006 shared task are available online. They have been published in (Buchholz and Marsi, 2006). The evaluation procedure was non-standard because it excluded punctuation tokens. These are the best results for Arabic:

The results of the CoNLL 2007 shared task are available online. They have been published in (Nivre et al., 2007). The evaluation procedure was changed to include punctuation tokens. These are the best results for Arabic:

The two Malt parser results of 2007 (single malt and blended) are described in (Hall et al., 2007) and the details about the parser configuration are described here.

BulTreeBank (BTB)

• Original BTB in native format
• CoNLL 2006 (BulTreeBank-DP)

The original BTB is based on HPSG (head-driven phrase-structure grammar). The CoNLL version contains only the dependency information encoded in HPSG BulTreeBank.

Only the CoNLL version seems to be distributed but you may ask the creators about the HPSG version. For the dependency version, print the license, sign, scan, send to Kiril Simov (kivs (at) bultreebank (dot) org) and wait for the data. The license in short:

• research usage
• no redistribution
• cite one publication in publications

BTB was created by members of the Linguistic Modelling Department (Секция Лингвистично моделиране), Bulgarian Academy of Sciences (Българска академия на науките), Ул. Акад. Г. Бончев, Бл. 25 А, 1113 София, Bulgaria.

Unknown (“A set of Bulgarian sentences marked-up with detailed syntactic information. These sentences are mainly extracted from authentic Bulgarian texts. They are chosen with regards two criteria. First, they cover the variety of syntactic structures of Bulgarian. Second, they show the statistical distribution of these phenomena in real texts.”) At least part of it is probably news (Novinar, Sega, Standart).

The CoNLL 2006 version contains 196,151 tokens in 13221 sentences, yielding 14.84 tokens per sentence on average (CoNLL 2006 data split: 190,217 tokens / 12823 sentences training, 5934 tokens / 398 sentences test).

• Website
  • http://www.bultreebank.org/indexBTB.html
• Data
  • no separate citation
• Principal publications
  • Kiril Simov, Petya Osenova, Alexander Simov, Milen Kouylekov: Design and Implementation of the Bulgarian HPSG-based Treebank. In: Erhard Hinrichs, Kiril Simov (eds.): Journal of Research on Language and Computation, Special Issue, vol. 2, no. 4, pp. 495–522, Kluwer Academic Publishers, ISSN 1570-7075. 2004.
• Documentation
  • Kiril Simov, Petya Osenova, Milena Slavcheva: BTB-TR03: BulTreeBank Morphosyntactic Tagset. Technical report, 2004.
  • Petya Osenova, Kiril Simov: BTB-TR05: BulTreeBank Stylebook. Technical report, 2004.
  • http://www.bultreebank.org/dpbtb/ provides the list of dependency relation labels (s-tags) with brief description.

The original morphosyntactic tags have been converted to fit into the three columns (CPOS, POS and FEAT) columns of the CoNLL format. There should be a 1-1 mapping between the BTB positional tags and the CoNLL 2006 annotation. Use DZ Interset to inspect the two CoNLL tagsets.

The morphological analysis does not include lemmas. The morphosyntactic tags have been assigned (probably) manually.

The guidelines for syntactic annotation are documented in the other technical report. The CoNLL distribution contains the BulTreeBankReadMe.html file with a brief description of the syntactic tags (dependency relation labels).

The first three sentences of the CoNLL 2006 training data:

The first three sentences of the CoNLL 2006 test data:

Nonprojectivities in BTB are rare. Only 747 of the 196,151 tokens in the CoNLL 2006 version are attached nonprojectively (0.38%).

The results of the CoNLL 2006 shared task are available online. They have been published in (Buchholz and Marsi, 2006). The evaluation procedure was non-standard because it excluded punctuation tokens. These are the best results for Bulgarian:

Hyderabad Dependency Treebank (HyDT-Bangla)

• ICON 2009
  • Shakti Standard Format (SSF; native)
  • CoNLL format
• ICON 2010
  • Shakti Standard Format (SSF; native)
  • CoNLL format

There has been no official release of the treebank yet. There have been two as-is sample releases for the purposes of the NLP tools contests in parsing Indian languages, attached to the ICON 2009 and 2010 conferences.

There is no standard distribution channel for the treebank after the ICON 2010 evaluation period. Inquire at the LTRC (ltrc (at) iiit (dot) ac (dot) in) about the possibility of getting the data. The ICON 2010 license in short:

• non-commercial research usage
• no redistribution

HyDT-Bangla is being created by members of the Language Technologies Research Centre, International Institute of Information Technology, Gachibowli, Hyderabad, 500032, India.

Unknown.

HyDT-Bangla shows dependencies between chunks, not words. The node/tree ratio is thus much lower than in other treebanks. The ICON 2010 version came with a data split into three parts: training, development and test:

I have counted the sentences and chunks. The number of words comes from (Husain et al., 2010). Note that the paper gives the number of training sentences as 980 (instead of 979), which is a mistake. The last training sentence has the id 980 but there is no sentence with id 418.

• Website
  • no website dedicated to the treebank
  • ICON 2009 NLP Tools Contest On-Line Proceedings
  • ICON 2010 NLP Tools Contest On-Line Proceedings
• Data
  • no separate citation
• Principal publications
  • Samar Husain: Dependency Parsers for Indian Languages. In: Proceedings of ICON-2009 NLP Tools Contest, Hyderabad, India, 2009.
  • Samar Husain, Prashanth Mannem, Bharat Ambati, Phani Gadde: The ICON-2010 tools contest on Indian language dependency parsing. In: Proceedings of ICON-2010 NLP Tools Contest, pp. 1–8, Kharagpur, India, 2010.
• Documentation
  • http://ltrc.iiit.ac.in/nlptools2010/documentation.php

The text uses the WX encoding of Indian letters. If we know what the original script is (Bengali in this case) we can map the WX encoding to the original characters in UTF-8. WX uses English letters so if there was embedded English (or other string using Latin letters) it will probably get lost during the conversion.

The CoNLL format contains only the chunk heads. The native SSF format shows the other words in the chunk, too, but it does not capture intra-chunk dependency relations. This is an example of a multi-word chunk:

3       ((      NP      <fs af='rumAla,n,,sg,,d,0,0' head="rumAla" drel=k2:VGF name=NP3>
3.1     ekatA   QC      <fs af='eka,num,,,,,,'>
3.2     ledisa  JJ      <fs af='ledisa,unk,,,,,,'>
3.3     rumAla  NN      <fs af='rumAla,n,,sg,,d,0,0' name="rumAla">
        ))

In the CoNLL format, the CPOS column contains the chunk label (e.g. NP = noun phrase) and the POS column contains the part of speech of the chunk head.

Occasionally there are NULL nodes that do not correspond to any surface chunk or token. They represent ellided participants.

The syntactic tags (dependency relation labels) are karaka relations, i.e. deep syntactic roles according to the Pāṇinian grammar. There are separate versions of the treebank with fine-grained and coarse-grained syntactic tags.

According to (Husain et al., 2010), in the ICON 2010 version, the chunk tags, POS tags and inter-chunk dependencies (topology + tags) were annotated manually. The rest (lemma, morphosyntactic features, headword of chunk) was marked automatically.

Note: There have been cycles in the Hindi part of HyDT but no such problem occurs in the Bengali part.

The first sentence of the ICON 2010 training data (with fine-grained syntactic tags) in the Shakti format:

<document id="">			
<head>
<annotated-resource name="HyDT-Bangla" version="0.5" type="dep-interchunk-only" layers="morph,pos,chunk,dep-interchunk-only" language="ben" date-of-release="20100831">
    <annotation-standard>
        <morph-standard name="Anncorra-morph" version="1.31" date="20080920" />
        <pos-standard name="Anncorra-pos" version="" date="20061215" />
        <chunk-standard name="Anncorra-chunk" version="" date="20061215" />
        <dependency-standard name="Anncorra-dep" version="2.0" date="" dep-tagset-granularity="6" />
    </annotation-standard>
</annotated-resource>		
</head>			
<Sentence id="1">
1	((	NP	<fs af='Age,adv,,,,,,' head="Agei" drel=k7t:VGF name=NP>
1.1	mudZira	NN	<fs af='mudZi,n,,sg,,o,era,era'>
1.2	Agei	NST	<fs af='Age,adv,,,,,,' name="Agei">
	))		
2	((	NP	<fs af='cA,n,,sg,,d,0,0' head="cA" drel=k1:VGF name=NP2>
2.1	praWama	QO	<fs af='praWama,num,,,,,,'>
2.2	kApa	NN	<fs af='kApa,unk,,,,,,'>
2.3	cA	NN	<fs af='cA,n,,sg,,d,0,0' name="cA">
	))		
3	((	VGF	<fs af='As,v,,,5,,A_yA+Ce,A' head="ese" name=VGF>
3.1	ese	VM	<fs af='As,v,,,7,,A,A' name="ese">
3.2	.	SYM	<fs af='.,punc,,,,,,'>
	))		
</Sentence>

And in the CoNLL format:

And after conversion of the WX encoding to the Bengali script in UTF-8:

The first sentence of the ICON 2010 development data (with fine-grained syntactic tags) in the Shakti format:

<document id="">
<head>
<annotated-resource name="HyDT-Bangla" version="0.5" type="dep-interchunk-only" layers="morph,pos,chunk,dep-interchunk-only" language="ben" date-of-release="20100831">
    <annotation-standard>
        <morph-standard name="Anncorra-morph" version="1.31" date="20080920" />
        <pos-standard name="Anncorra-pos" version="" date="20061215" />
        <chunk-standard name="Anncorra-chunk" version="" date="20061215" />
        <dependency-standard name="Anncorra-dep" version="2.0" date="" dep-tagset-granularity="6" />
    </annotation-standard>
</annotated-resource>
</head>
<Sentence id="1">
1	((	NP	<fs af='parabarwIkAle,adv,,,,,,' head="parabarwIkAle" drel=k7t:VGF name=NP>
1.1	parabarwIkAle	NN	<fs af='parabarwIkAle,adv,,,,,,' name="parabarwIkAle">
	))		
2	((	NP	<fs af='aPisa-biyArAraxera,unk,,,,,,' head="aPisa-biyArAraxera" drel=r6:NP3 name=NP2>
2.1	aPisa-biyArAraxera	NN	<fs af='aPisa-biyArAraxera,unk,,,,,,' name="aPisa-biyArAraxera">
	))		
3	((	NP	<fs af='nAma,n,,sg,,d,0,0' head="nAma" drel=k2:VGNN name=NP3>
3.1	nAma	NN	<fs af='nAma,n,,sg,,d,0,0' name="nAma">
	))		
4	((	NP	<fs af='GoRaNA,unk,,,,,,' head="GoRaNA" drel=pof:VGNN name=NP4>
4.1	GoRaNA	NN	<fs af='GoRaNA,unk,,,,,,' name="GoRaNA">
	))		
5	((	VGNN	<fs af='kar,n,,,any,,,' head="karAra" drel=r6:NP5 name=VGNN>
5.1	karAra	VM	<fs af='kar,n,,,any,,,' name="karAra">
	))		
6	((	NP	<fs af='samay,unk,,,,,,' head="samay" drel=k7t:VGF name=NP5>
6.1	samay	NN	<fs af='samay,unk,,,,,,' name="samay">
	))		
7	((	NP	<fs af='animeRake,unk,,,,,,' head="animeRake" drel=k2:VGF name=NP6>
7.1	animeRake	NNP	<fs af='animeRake,unk,,,,,,' name="animeRake">
	))		
8	((	VGF	<fs af='sariye,unk,,,5,,0_rAKA+ka_ha+la,' head="sariye" name=VGF>
8.1	sariye	VM	<fs af='sariye,unk,,,,,,' name="sariye">
8.2	.	SYM	<fs af='.,punc,,,,,,'>
	))		
</Sentence>

And in the CoNLL format:

And after conversion of the WX encoding to the Bengali script in UTF-8:

The first sentence of the ICON 2010 test data (with fine-grained syntactic tags) in the Shakti format:

<document id="">
<head>
<annotated-resource name="HyDT-Bangla" version="0.5" type="dep-interchunk-only" layers="morph,pos,chunk,dep-interchunk-only" language="ben" date-of-release="20101013">
    <annotation-standard>
        <morph-standard name="Anncorra-morph" version="1.31" date="20080920" />
	<pos-standard name="Anncorra-pos" version="" date="20061215" />
	<chunk-standard name="Anncorra-chunk" version="" date="20061215" />
	<dependency-standard name="Anncorra-dep" version="2.0" date="" dep-tagset-granularity="6" />
    </annotation-standard>
<annotated-resource>
</head>
<Sentence id="1">
1	((	NP	<fs af='mAXabIlawA,n,,sg,,d,0,0' head="mAXabIlawA" drel=k1:VGF name=NP>
1.1	mAXabIlawA	NNP	<fs af='mAXabIlawA,n,,sg,,d,0,0' name="mAXabIlawA">
	))		
2	((	NP	<fs af='waKana,pn,,,,d,0,0' head="waKana" drel=k7t:VGF name=NP2>
2.1	waKana	PRP	<fs af='waKana,pn,,,,d,0,0' name="waKana">
	))		
3	((	NP	<fs af='hAwa,n,,sg,,o,era,era' head="hAwera" drel=r6:NP4 name=NP3>
3.1	hAwera	NN	<fs af='hAwa,n,,sg,,o,era,era' name="hAwera">
	))		
4	((	NP	<fs af='GadZi,unk,,,,,,' head="GadZi" drel=k2:VGNF name=NP4>
4.1	GadZi	NN	<fs af='GadZi,unk,,,,,,' name="GadZi">
	))		
5	((	VGNF	<fs af='Kul,v,,,5,,ne,ne' head="Kule" drel=vmod:VGF name=VGNF>
5.1	Kule	VM	<fs af='Kul,v,,,5,,ne,ne' name="Kule">
	))		
6	((	NP	<fs af='tebila,n,,sg,,d,me,me' head="tebile" drel=k7p:VGF name=NP5>
6.1	tebile	NN	<fs af='tebila,n,,sg,,d,me,me' name="tebile">
	))		
7	((	VGF	<fs af='rAK,v,,,5,,Cila,Cila' head="rAKaCila" name=VGF>
7.1	rAKaCila	VM	<fs af='rAK,v,,,5,,Cila,Cila' name="rAKaCila">
7.2	।	SYM	
	))		
</Sentence>

And in the CoNLL format:

And after conversion of the WX encoding to the Bengali script in UTF-8: