TL;DR
This paper introduces the first open-access parallel corpus for Spanish and Chinese, annotated with discourse structure information based on Rhetorical Structure Theory (RST). This resource aims to bridge a critical gap in Natural Language Processing (NLP) research by enabling comparative discourse analysis between these two widely spoken languages, paving the way for improved machine translation and cross-lingual understanding.
What Problem Does This Paper Address?
Natural Language Processing (NLP) tasks, such as text summarization, question answering, and machine translation, increasingly rely on understanding the underlying discourse structure of text. However, research in this area has been hampered by a lack of comprehensive, parallel corpora for many language pairs, particularly for languages with significant structural differences like Spanish and Chinese. Existing discourse analysis tools and theories, often developed for English, struggle to capture the nuances of other languages. This paper addresses the need for a robust, annotated resource that facilitates the study of discourse similarities and differences between Spanish and Chinese, enabling the development of more effective cross-lingual NLP applications.
Methods
The authors developed a Spanish-Chinese parallel corpus annotated according to Rhetorical Structure Theory (RST). The methodology involved several key steps:
- Segmentation: They established specific criteria for segmenting texts into Elementary Discourse Units (EDUs), considering both linguistic function and form (punctuation, verbs). This process was carried out by bilingual and monolingual linguists.
- Central Unit (CU) Annotation: For each segmented text, annotators identified the Central Unit (CU), representing the core information of the text, following established RST practices.
- Discourse Structure Annotation: The core of the work involved annotating the rhetorical relations between EDUs using the RST framework. This was done using the RSTTool and rstWeb.
- Harmonization: To facilitate cross-lingual comparison, a harmonization process was employed, particularly for discourse structure annotation, where Chinese EDUs were sometimes combined to match the Spanish structure, ensuring comparable units for analysis.
- Evaluation: Inter-annotator agreement was measured using Cohen's Kappa for both segmentation and CU annotation, and a qualitative method based on F-measure for Nuclearity, Relation, Composition, and Attachment in discourse structure annotation.
Datasets
The paper introduces a newly created Spanish-Chinese parallel corpus. This corpus consists of 100 texts, with 50 Spanish texts and their corresponding Chinese translations. The texts were selected to represent a variety of sizes (111-1774 words), genres (abstracts, news, advertisements, announcements), domains (terminology, culture, language, economy, education, art, international affairs), and authors, aiming for complexity and heterogeneity in discourse structure. This corpus is made publicly available.
Key Insights and Ideas
- Cross-lingual Discourse Differences are Significant: The paper highlights that Spanish and Chinese exhibit substantial differences in their discourse structures, even when conveying similar meanings. For instance, the use and order of discourse markers can vary considerably, impacting sentence structure and grammatical requirements.
- RST as a Unifying Framework: Rhetorical Structure Theory provides a robust theoretical framework that can be applied to analyze and compare discourse structures across languages, despite their inherent differences.
- Importance of Bilingual Annotators: The success of the annotation process relied heavily on the expertise of bilingual annotators who could navigate the linguistic nuances of both Spanish and Chinese.
- Harmonization is Crucial for Cross-Lingual Comparison: To effectively compare discourse structures between languages, a harmonization step is necessary to align the units of analysis, even if it means slightly altering the original segmentation to achieve comparability.
Key Results
- High Annotation Quality: The evaluation results demonstrate high inter-annotator agreement for both segmentation (Spanish: 0.87, Chinese: 0.76) and Central Unit annotation (Spanish: 0.961, Chinese: 0.977), indicating the reliability of the annotation criteria and process.
- Reliable Discourse Structure Annotation: Qualitative evaluation of the harmonized corpus showed strong agreement across different RST components (Nuclearity, Relation, Composition, Attachment) for both languages, with scores generally above 0.85.
- Corpus Availability: The developed Spanish-Chinese RST Treebank is made publicly available online, including an interface for searching POS information, discourse segments, CUs, and discourse structures, as well as downloading the texts.
Limitations and Future Work
- Limited Corpus Size: The authors acknowledge that due to resource constraints, the corpus size (100 texts) is limited, although it was chosen to ensure complexity and heterogeneity.
- Harmonization Trade-offs: The harmonization process, while necessary for comparison, can sometimes lead to the loss of some rhetorical relations at a lower level, though higher-level RS-Tree structures are generally preserved.
- Future Work: The authors plan to expand the corpus by selecting and annotating more Spanish-Chinese parallel texts. They also aim to develop a protocol to assist machine translation (MT) for this language pair, leveraging the discourse information captured in the corpus.
Summary updated