Kyoto University Web Document Leads Corpus

Overview

This is a Japanese text corpus that consists of lead three sentences of web documents with various linguistic annotations. By collecting lead three sentences of web documents, this corpus contains documents with various genres and styles, such as news articles, encyclopedic articles, blogs and commercial pages. It comprises approximately 5,000 documents, which correspond to 15,000 sentences.

The linguistic annotations consist of annotations of morphology, named entities, dependencies, predicate-argument structures including zero anaphora, coreferences, and discourse. All the annotations except discourse annotations were given by manually modifying automatic analyses of the morphological analyzer JUMAN and the dependency, case structure and anaphora analyzer KNP. The discourse annotations were given using crowdsourcing.

Notes

This corpus consists of linguistically annotated Web documents that have been made publicly available on the Web at some time. The corpus is released for the purpose of contributing to the research of natural language processing.

Since the collected documents are fragmentary, i.e., only the lead three sentences of each Web document, we have not obtained permission from copyright owners of the Web documents and do not provide source information such as URL. If copyright owners of Web documents request addition of source information or deletion of these documents, we will update the corpus and newly release it. In this case, please delete the downloaded old version and replace it with the new version.

Notes on annotation guidelines

The annotation guidelines for this corpus are written in the manuals found in "doc" directory. The guidelines for morphology and dependencies are described in syn_guideline.pdf, those for predicate-argument structures and coreferences are described in rel_guideline.pdf, and those for discourse relations are described in disc_guideline.pdf. The guidelines for named entities are available at the IREX web site (http://nlp.cs.nyu.edu/irex/).

Distributed files

• knp/ : the corpus annotated with morphology, named entities, dependencies, predicate-argument structures, and coreferences
• disc/ : the corpus annotated with discourse relations
• org/ : the raw corpus
• doc/ : annotation guidelines
• id/ : document id files providing train/test split

Note that the encoding of the corpus data is UTF-8.

Format of the corpus annotated with annotations of morphology, named entities, dependencies, predicate-argument structures, and coreferences

Annotations of this corpus are given in the following format.

# S-ID:w201106-0000010001-1
* 2D
+ 3D
太郎 たろう 太郎 名詞 6 人名 5 * 0 * 0
は は は 助詞 9 副助詞 2 * 0 * 0
* 2D
+ 2D
京都 きょうと 京都 名詞 6 地名 4 * 0 * 0
+ 3D <NE:ORGANIZATION:京都大学>
大学 だいがく 大学 名詞 6 普通名詞 1 * 0 * 0
に に に 助詞 9 格助詞 1 * 0 * 0
* -1D
+ -1D <rel type="ガ" target="太郎" sid="w201106-0000010001-1" id="0"/><rel type="ニ" target="大学" sid="w201106-0000010001-1" id="2"/>
行った いった 行く 動詞 2 * 0 子音動詞カ行促音便形 3 タ形 10
EOS

The first line represents the ID of this sentence. In the subsequent lines, the lines starting with "*" denote "bunsetsu," the lines starting with "+" denote basic phrases, and the other lines denote morphemes.

The line of morphemes is the same as the output of the morphological analyzers, JUMAN and Juman++. This information includes surface string, reading, lemma, part of speech (POS), fine-grained POS, conjugate type, and conjugate form. "*" means that its field is not available. Note that this format is slightly different from KWDLC 1.0, which adopted the same format as Kyoto University Text Corpus 4.0.

The line starting with "*" represents "bunsetsu," which is a conventional unit for dependency in Japanese. "Bunsetsu" consists of one or more content words and zero or more function words. In this line, the first numeral means the ID of its depending head. The subsequent alphabet denotes the type of dependency relation, i.e., "D" (normal dependency), "P" (coordination dependency), "I" (incomplete coordination dependency), and "A" (appositive dependency).

The line starting with "+" represents a basic phrase, which is a unit to which various relations are annotated. A basic phrase consists of one content word and zero or more function words. Therefore, it is equivalent to a bunsetsu or a part of a bunsetsu. In this line, the first numeral means the ID of its depending head. The subsequent alphabet is defined in the same way as bunsetsu. The remaining part of this line includes the annotations of named entity and various relations.

Annotations of named entity are given in <NE> tags. <NE> has the following four attributes: type, target, possibility, and optional_type, which mean the class of a named entity, the string of a named entity, possible classes for an OPTIONAL named entity, and a type for an OPTIONAL named entity, respectively. The details of these attributes are described in the IREX annotation guidelines.

Annotations of various relations are given in <rel> tags. <rel> has the following four attributes: type, target, sid, and id, which mean the name of a relation, the string of the counterpart, the sentence ID of the counterpart, and the basic phrase ID of the counterpart, respectively. If a basic phrase has multiple tags of the same type, a "mode" attribute is also assigned, which has one of "AND," "OR," and "？." The details of these attributes are described in the annotation guidelines (rel_guideline.pdf).

Format of the corpus annotated with discourse relations

In this corpus, a clause pair is given a discourse type and its probability as follows.

# A-ID:w201106-0001998536
1 今日とある企業のトップの話を聞くことが出来た。
2 経営者として何事も全てビジネスチャンスに変えるマインドが大切だと感じた。
3 生きていく上で追い風もあれば、
4 逆風もある。
1-2 関係なしまたは弱い関係:0.999915 対比:3.6e-05 根拠:1.5e-05 原因・理由:8e-06 目的:7e-06
3-4 対比:0.999986 その他根拠:3e-06

The first line represents the ID of this document, the subsequent block denotes clause IDs and clauses, and the last block denotes discourse relations for clause pairs and their probabilities. These discourse relations and probabilities are the results of the second stage of crowdsourcing. Each line is the list of a discourse relation and its probability in order of probability. For the discourse relation with the highest probability, the discourse direction is annotated; if it is reverse order, "(逆方向)" is added to the discourse relation. The details of these probabilities and discourse relations are described in [Kawahara et al., 2014] and the annotation guidelines (disc_guideline.pdf).

References

• Masatsugu Hangyo, Daisuke Kawahara and Sadao Kurohashi. Building a Diverse Document Leads Corpus Annotated with Semantic Relations, In Proceedings of the 26th Pacific Asia Conference on Language Information and Computing, pp.535-544, 2012. http://www.aclweb.org/anthology/Y/Y12/Y12-1058.pdf
• 萩行正嗣, 河原大輔, 黒橋禎夫. 多様な文書の書き始めに対する意味関係タグ付きコーパスの構築とその分析, 自然言語処理, Vol.21, No.2, pp.213-248, 2014. https://doi.org/10.5715/jnlp.21.213
• Daisuke Kawahara, Yuichiro Machida, Tomohide Shibata, Sadao Kurohashi, Hayato Kobayashi and Manabu Sassano. Rapid Development of a Corpus with Discourse Annotations using Two-stage Crowdsourcing, In Proceedings of the 25th International Conference on Computational Linguistics, pp.269-278, 2014. http://www.aclweb.org/anthology/C/C14/C14-1027.pdf

Acknowledgment

The creation of this corpus was supported by JSPS KAKENHI Grant Number 24300053 and JST CREST "Advanced Core Technologies for Big Data Integration." The discourse annotations were acquired by crowdsourcing under the support of Yahoo! Japan Corporation. We deeply appreciate their support.

Contact

If you have any questions or problems about this corpus, please send an email to nl-resource at nlp.ist.i.kyoto-u.ac.jp. If you have a request to add source information or to delete a document in the corpus, please send an email to this mail address.