MCDTB: A Macro-level Chinese Discourse TreeBank

Proceedings of the 27th International Conference on Computational Linguistics, pages 3493–3504 Santa Fe, New Mexico, USA, August 20-26, 2018. 3493 MCDTB: A Macro-Level Chinese Discourse TreeBank Feng Jiang1, Sheng Xu1, Xiaomin Chu1, Peifeng Li1, Qiaoming Zhu12, Guodong Zhou1 1School of Computer Science and Technology, Soochow University, China 2Institute of Artificial Intelligence, Soochow University, China {fjiang, sxu, xmchu}@stu.suda.edu.cn {pfli, qmzhu, gdzhou}@suda.edu.cn Abstract In view of the differences between the annotations of micro and macro discourse relationships, this paper describes the relevant experiments on the construction of the Macro Chinese Dis- course Treebank (MCDTB), a higher-level Chinese discourse corpus. Following RST (Rhetori- cal Structure Theory), we annotate the macro discourse information, including discourse struc- ture, nuclearity and relationship, and the additional discourse information, including topic sen- tences, lead and abstract, to make the macro discourse annotation more objective and accurate. Finally, we annotated 720 articles with a Kappa value greater than 0.6. Preliminary experiments on this corpus verify the computability of MCDTB. 1 Introduction In the field of natural language processing, discourse analysis is becoming increasingly important as the object of research gradually shifts from the word level to sentence, event and other semantic aspects. Discourse analysis primarily examines the text coherence and cohesion, including the analysis on struc- ture, nuclearity and relationship. In discourse analysis, discourse refers to a series of continuous clauses, sentences or paragraphs of language as a whole; it not only includes text sequences but also the text structure and the logical relationship between text sequences. Discourse analysis aims at studying the internal structure of texts and understanding the semantic relation between different text units. The granularity of this research can be clause, sentence, sentence group, paragraph and whole article. Commonly, discourse analysis is divided into microstructure and macrostructure analysis. The former is the study of intra-sentence or inter-sentence discourse relations, while the latter is discourse relation between sentence clusters, paragraphs and chapters (Van Dijk, 1976), which highlights a higher semantic level for text comprehension. Example 1 is the content of the article chtb0155, which has a title (T) and five paragraphs (P1-P5), i.e., five discourse units. Figure 1 shows the example’s macro discourse structure tree. The leaf nodes of the macro discourse structure tree refer to discourse units, and the internal nodes refer to relational nodes, which represent the larger discourse units of the relevant children. When connecting the parent and child nodes, the directed edge indicates that the child is an important discourse unit in the relation- ship, and the undirected edge indicates that the child is secondary in the relationship. Therefore, in Fig- ure 1, there is an Evaluation relation between the discourse units P1 and P2, and they form a higher level unit (relational node) DU1-2. The discourse units P4 and P5 exhibit the Purpose-Behaviour relation and form a higher level unit DU4-5. Consequently, DU1-2 and P3 have the Elaboration relation and form a higher level unit DU1-3. Finally, DU4-5 is supplementary to DU1-3. Based on the discourse structure tree shown in Figure 1, …

Content truncated to 500 words