Repository files navigation

JaLeCoN is a Dataset ofJapaneseLexicalComplexity forNon-Native Readers. It can be used to train or evaluate Japanese lexical complexity prediction models.

Please cite the followingpaper in publications using this dataset. You can also refer to it for details about the data and a baseline experiment.

@inproceedings{ide2023,  title     = "Japanese Lexical Complexity for Non-Native Readers: A New Dataset",  booktitle = "Proceedings of the Eighteenth Workshop on Innovative Use of {NLP} for Building Educational Applications",  author    = "Ide, Yusuke and Mita, Masato and Nohejl, Adam and Ouchi, Hiroki and Watanabe, Taro",  month     = July,  year      = 2023,  publisher = "Association for Computational Linguistics",}

ck.txt,non_ck.txt

• These files contain data for each L1 group. See alsoData Format below.

cv_splits.json

• This file contains cross-validation splits. See alsoCross-Validation Splits below.

mwe_list

• This directory contains manually collected lists of ordinary MWEs inmwe.tsv and proper noun MWEs inproper_mwe.tsv. Sequences of SUWs that match any MWEs in the lists are chunked in the step of MWE annotation (see Section 3.3 ofour paper).| in the MWEs denotes a boundary between SUWs.

The filesck.txt andnon_ck.txt are composed of sentences, each of which has the following format (terminated by a blank line):

<ID>\t<sentence_num>\t<sentence><word_1>[…]<word_n>

Example:

news-1-11ネット|情報|から|危機|発生|を|瞬時|に|検知|、|デマ|も|判定|（|JBpress|）|-|Yahoo|!|ニュースネット0:1-70000.0情報1:2-61000.047[…]-16:17-70000.0Yahoo!ニュース17:20P-----

The first line is in the format<ID>\t<sentence_num>\t<sentence>:

• ID is the unique ID of the sentence in the format<genre>-<block>-<line>, wheregenre identifies the genre (news orgov);block is a unit annotated by the same set of annotators;line is the line number in the block.
• sentence_num is the sentence number in the sequence to which the sentence belongs. This is useful for knowing where each sequence starts and ends.
• sentence is the sentence itself segmented into SUWs (delimited with|).

The subsequent<word_i> lines provide information about each word in the sentence in the format<word>\t<position>\t<ignore>\t<easy>\t<not_easy>\t<diff>\t<very_diff>\t<complexity>:

• position consists of the start and end offsets of the word separated with a colon.
• ignore containsP orQ when the word should be ignored. Specifically,P means that the word is a proper noun;S means that an error related to segmentation occurred in the word.
• easy,not_easy,diff, andvery_diff are the numbers of annotators who assigned the respective label to the word.
• complexity is the complexity score of the word for the L1 group.

When the word is to be ignored, the fieldseasy,not_easy,diff,very_diff, andcomplexity each contain-.

The filecv_splits.json contains cross-validation splits that we have used for the experiment inour paper, and which can therefore be used to replicate it or to easily compare our results with other systems.

We have used:

• 5 folds,
• stratification by genre (News and Government),
• grouping by sequence of sentences (see Section 3 inthe paper).

Genres are thus equally represented in each fold, and a sequence of sentences is never split between a training set and a validation set.

cv_splits.json has the following form:

[[split1_train, split1_validation],[split2_train, split2_validation],[split3_train, split3_validation],[split4_train, split4_validation],[split5_train, split5_validation]]

Each split's training or validation set is a list of IDs (first tab-separated fields of the sentence lines in the dataset files).

We used data from the Japanese-English development set ofthe WMT22 General Machine Translation Task (Kocmi et al. 2022¹).

Website of the Japan Meteorological Agency (気象庁ウェブサイト):

https://www.jma.go.jp/jma/kishou/tyoukan/2023/dg_20230105.htmlhttps://www.jma.go.jp/jma/kishou/tyoukan/2022/dg_20221221.htmlhttps://www.jma.go.jp/jma/kishou/tyoukan/2022/dg_20221116.htmlhttps://www.jma.go.jp/jma/kishou/tyoukan/2022/dg_20221018.html

Website of the Japan Tourism Agency (観光庁ウェブサイト):

https://www.mlit.go.jp/kankocho/page01_000728.htmlhttps://www.mlit.go.jp/kankocho/page01_000723.htmlhttps://www.mlit.go.jp/kankocho/page01_000718.htmlhttps://www.mlit.go.jp/kankocho/page01_000715.htmlhttps://www.mlit.go.jp/kankocho/page01_000712.html

Website of the Ministry of Justice (法務省ウェブサイト):

https://www.moj.go.jp/hisho/kouhou/hisho08_00386.htmlhttps://www.moj.go.jp/hisho/kouhou/hisho08_00381.htmlhttps://www.moj.go.jp/hisho/kouhou/hisho08_00380.htmlhttps://www.moj.go.jp/hisho/kouhou/hisho08_00375.htmlhttps://www.moj.go.jp/hisho/kouhou/hisho08_00344.htmlhttps://www.moj.go.jp/hisho/kouhou/hisho08_00341.html

Website of the Ministry of Foreign Affairs (外務省ウェブサイト):

https://www.mofa.go.jp/mofaj/press/kaiken/kaiken22_000066.htmlhttps://www.mofa.go.jp/mofaj/press/kaiken/kaiken24_000174.htmlhttps://www.mofa.go.jp/mofaj/press/kaiken/kaiken4_001075.htmlhttps://www.mofa.go.jp/mofaj/press/kaiken/kaiken24_000165.htmlhttps://www.mofa.go.jp/mofaj/press/kaiken/kaiken24_000164.html

Website of the Ministry of Health, Labour and Welfare (厚生労働省ウェブサイト):

https://www.mhlw.go.jp/stf/kaiken/daijin/0000194708_00527.htmlhttps://www.mhlw.go.jp/stf/kaiken/daijin/0000194708_00525.htmlhttps://www.mhlw.go.jp/stf/kaiken/daijin/0000194708_00522.htmlhttps://www.mhlw.go.jp/stf/kaiken/daijin/0000194708_00516.htmlhttps://www.mhlw.go.jp/stf/kaiken/daijin/0000194708_00515.htmlhttps://www.mhlw.go.jp/stf/kaiken/daijin/0000194708_00514.html

This work is licensed under aCreative Commons Attribution-NonCommercial-ShareAlike 4.0 International License.