GCDT: Georgetown Chinese Discourse Treebank

GCDT is the largest (as of October 2022) hierarchical discourse treebank for Mandarin Chinese in the framework of Rhetorical Structure Theory (RST).

GCDT covers over 60K tokens across five genres of freely available text, using the same relation inventory as the contemporary RST treebank for English -- GUM (https://github.com/amir-zeldes/gum).

Corpus statistics

The corpus contains 50 documents, 10 from each of the five genres below.

In total, there are 62,905 tokens and 9,717 EDUs.

We provide an 8:1:1 train:dev:test split for each genre as shown in the table below.

The five test documents are double annotated and the second annotation version appears in the double/ partition.

Genre	#docs	Avg #tokens/doc	Avg #EDUs/doc	Source	Dev doc	Test doc
academic	10	1416.8	203.3	https://www.hanspub.org/	gcdt_academic_peoples	gcdt_academic_dingzhen
bio	10	1348.5	202.1	https://zh.wikipedia.org/	gcdt_bio_byron	gcdt_bio_dvorak
interview	10	1146.4	181.2	https://zh.wikinews.org/	gcdt_interview_ward	gcdt_interview_wimax
news	10	1124.9	165.2	https://zh.wikinews.org/	gcdt_news_famine	gcdt_news_simplified
wikihow	10	1253.9	219.9	https://zh.wikihow.com/	gcdt_whow_hiking	gcdt_whow_thanksgiving

Citing

Please cite the following for the source paper -- GCDT: A Chinese RST Treebank for Multigenre and Multilingual Discourse Parsing:

@inproceedings{peng_gcdt_2022,
	address = {Online only},
	title = {{GCDT}: {A} {Chinese} {RST} {Treebank} for {Multigenre} and {Multilingual} {Discourse} {Parsing}},
	shorttitle = {{GCDT}},
	url = {https://aclanthology.org/2022.aacl-short.47},
	abstract = {A lack of large-scale human-annotated data has hampered the hierarchical discourse parsing of Chinese. In this paper, we present GCDT, the largest hierarchical discourse treebank for Mandarin Chinese in the framework of Rhetorical Structure Theory (RST). GCDT covers over 60K tokens across five genres of freely available text, using the same relation inventory as contemporary RST treebanks for English. We also report on this dataset's parsing experiments, including state-of-the-art (SOTA) scores for Chinese RST parsing and RST parsing on the English GUM dataset, using cross-lingual training in Chinese and English with multilingual embeddings.},
	urldate = {2022-11-22},
	booktitle = {Proceedings of the 2nd {Conference} of the {Asia}-{Pacific} {Chapter} of the {Association} for {Computational} {Linguistics} and the 12th {International} {Joint} {Conference} on {Natural} {Language} {Processing}},
	publisher = {Association for Computational Linguistics},
	author = {Peng, Siyao and Liu, Yang Janet and Zeldes, Amir},
	month = nov,
	year = {2022},
	pages = {382--391},
	file = {Full Text PDF:/Users/loganpeng/Zotero/storage/IEPKWVJH/Peng et al. - 2022 - GCDT A Chinese RST Treebank for Multigenre and Mu.pdf:application/pdf},
}

Please cite the following for the Chinese Discourse Annotation Reference Manual:

@techreport{peng_chinese_2022,
	type = {Research {Report}},
	title = {Chinese {Discourse} {Annotation} {Reference} {Manual}},
	url = {https://hal.archives-ouvertes.fr/hal-03821884},
	abstract = {This document provides extensive guidelines and examples for Rhetorical Structure Theory (RST) annotation in Mandarin Chinese. The guideline is divided into three sections. We first introduce preprocessing steps to prepare data for RST annotation. Secondly, we discuss syntactic criteria to segment texts into Elementary Discourse Units (EDUs). Lastly, we provide examples to define and distinguish discourse relations in different genres. We hope that this reference manual can facilitate RST annotations in Chinese and accelerate the development of the RST framework across languages.},
	urldate = {2022-11-30},
	institution = {Georgetown University (Washington, D.C.)},
	author = {Peng, Siyao and Liu, Yang Janet and Zeldes, Amir},
	month = oct,
	year = {2022},
	keywords = {Chinese, Discourse Analysis Representation, Rhetorical Structure Theory RST},
	file = {HAL PDF Full Text:/Users/loganpeng/Zotero/storage/Q7ZFHQ3Q/Peng et al. - 2022 - Chinese Discourse Annotation Reference Manual.pdf:application/pdf},
}

Data, utils and adapted DMRST parser

Please see the sub-directories for annotated data, scripts, and adapted DMRST parser.

• data/ https://github.com/logan-siyao-peng/GCDT/tree/main/data
• utils/ https://github.com/logan-siyao-peng/GCDT/tree/main/utils
• adapted_DMRST/ https://github.com/logan-siyao-peng/GCDT/tree/main/adapted_DMRST

Data and metadata annotations

• XML and metadata annotations -- gold
• Paragraph and sentence splits -- gold
• Tokenization -- gold
• Dependency parses -- predicted by stanza
• rs3 to rsd, dis conversions -- automatic
• EDU-wise Chinese --> English translation -- automatic

Relation distributions in GCDT

We present GCDT's distribution of the 32 discourse relations in the table below.

relation name	total counts	percentage
joint-list	2808	22.16
same-unit	2375	18.74
elaboration-attribute	992	7.83
joint-sequence	647	5.10
joint-other	595	4.69
elaboration-additional	551	4.35
attribution-positive	551	4.35
explanation-evidence	522	4.12
adversative-contrast	412	3.25
context-circumstance	350	2.76
context-background	350	2.76
organization-preparation	288	2.27
causal-cause	238	1.88
contingency-condition	228	1.80
organization-heading	225	1.78
adversative-concession	208	1.64
purpose-goal	193	1.52
restatement-partial	164	1.29
evaluation-comment	136	1.07
mode-means	132	1.04
causal-result	123	0.97
explanation-justify	120	0.95
joint-disjunction	92	0.73
adversative-antithesis	78	0.62
mode-manner	75	0.59
topic-question	57	0.45
restatement-repetition	40	0.32
explanation-motivation	35	0.28
organization-phatic	33	0.26
attribution-negative	29	0.23
purpose-attribute	26	0.21
topic-solutionhood	1	0.01