Table of Contents

Japanese (ja)

Tübingen Treebank of Spoken Japanese (TüBa-J/S, Verbmobil project)

Versions

The original TüBa-J/S is HPSG-oriented, there is the lexical level, the phrasal level, the clausal level, and dependency edges between nodes. The CoNLL version contains only the dependency relations.

Obtaining and License

To obtain the treebank, download the license agreement, print it, fill it out and sign it, scan and send it back to Kathrin Beck (kbeck (at) sfs (dot) uni-tuebingen (dot) de). She will send you the password for the download page. The license in short:

TüBa-J/S was created in the Verbmobil project by members of the Seminar für Sprachwissenschaft, Eberhard Karls Universität Tübingen, Wilhelmstrasse 19, D-72074 Tübingen, Germany.

References

Domain

Spoken dialogues, negotiations about time and place of business meetings. That is why many sentences are relatively short (a frequent single-word sentence is hai = “yes”).

Size

The CoNLL 2006 version contains 157,172 tokens in 17753 sentences, yielding 8.85 tokens per sentence on average (CoNLL 2006 data split: 151,461 tokens / 17044 sentences training, 5711 tokens / 709 sentences test).

Inside

The text has been romanized and the original characters (kanji + kana) are not available. There should be a 1-1 mapping between the romanized text (rōmaji) and the Japanese script of hiragana. There is no indication though where katakana or kanji are preferred over hiragana.

The morphological analysis does not include lemmas. The part-of-speech tags have been assigned (probably) manually. Few morphosyntactic features are used.

Sample

The first three sentences of the CoNLL 2006 training data:

1 kasahara _ NAME NAMEper _ 2 HD 2 HD
2 arisa _ NAME NAMEper _ 3 COMP 3 COMP
3 desu _ PV PVfin u 0 ROOT 0 ROOT
4 . _ . . _ 3 PUNCT 3 PUNCT
1 kadowaki _ NAME NAMEper _ 2 HD 2 HD
2 masakazu _ NAME NAMEper _ 3 COMP 3 COMP
3 desu _ PV PVfin u 0 ROOT 0 ROOT
4 . _ . . _ 3 PUNCT 3 PUNCT
1 kadowaki _ NAME NAMEper _ 2 COMP 2 COMP
2 saN _ P PNsf _ 0 ROOT 0 ROOT
3 omatase _ N VN _ 4 COMP 4 COMP
4 shimashita _ VS VSfin ta 0 ROOT 0 ROOT
5 . _ . . _ 4 PUNCT 4 PUNCT

The first three sentences of the CoNLL 2006 test data:

1 tashiro _ NAME NAMEper _ 2 HD 2 HD
2 yasuko _ NAME NAMEper _ 3 COMP 3 COMP
3 desu _ PV PVfin u 0 ROOT 0 ROOT
4 . _ . . _ 3 PUNCT 3 PUNCT
1 hayakawa _ NAME NAMEper _ 2 HD 2 HD
2 ryou _ NAME NAMEper _ 3 COMP 3 COMP
3 desu _ PV PVfin u 0 ROOT 0 ROOT
4 . _ . . _ 3 PUNCT 3 PUNCT
1 hayakawa _ NAME NAMEper _ 2 COMP 2 COMP
2 saN _ P PNsf _ 6 ADJ 6 ADJ
3 ima _ N Ntmp _ 6 ADJ 6 ADJ
4 chotto _ ADV ADV _ 6 ADJ 6 ADJ
5 ojikaN _ N NN _ 6 SBJ 6 SBJ
6 yoroshii _ ADJ ADJifin _ 7 COMP 7 COMP
7 desu _ PV PVfin u 0 ROOT 0 ROOT
8 ka _ PS PSE _ 7 MRK 7 MRK
9 chotto _ ADV ADV _ 16 ADJ 16 ADJ
10 doitsu _ NAME NAMEloc _ 11 COMP 11 COMP
11 no _ P Pgen _ 12 ADJ 12 ADJ
12 shucchou _ N VN _ 13 COMP 13 COMP
13 no _ P Pgen _ 14 COMP 14 COMP
14 koto _ N NF _ 15 COMP 15 COMP
15 de _ P P _ 16 ADJ 16 ADJ
16 gosoudaN _ N VN _ 17 COMP 17 COMP
17 shitai _ VADJ VADJi _ 18 COMP 18 COMP
18 no _ N NF _ 19 COMP 19 COMP
19 desu _ PV PVfin u 20 COMP 20 COMP
20 ga _ PS PSSb _ 0 ROOT 0 ROOT
21 . _ . . _ 20 PUNCT 20 PUNCT

Parsing

Nonprojectivities in TüBa-J/S are not frequent. Only 1736 of the 157,172 tokens in the CoNLL 2006 version are attached nonprojectively (1.1%).

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 Japanese:

Parser (Authors) LAS UAS
Basis (John O'Neil) 90.57 93.16
Nara (Yuchang Cheng) 89.91 93.12
Malt (Nivre et al.) 91.65 93.10