An Annotation System for Development of Chinese Discourse Corpus

Hen-Hsen Huang, Hsin-Hsi ChenCOLING 2012

TL;DR

This paper introduces a web-based annotation system designed to create a fine-grained discourse corpus for Chinese. By addressing the unique linguistic characteristics of Chinese and incorporating sentiment analysis, this work aims to facilitate deeper research into Chinese discourse structure and its relationship with sentiment.

What Problem Does This Paper Address?

Discourse analysis, which studies how sentences and clauses connect to form coherent texts, relies heavily on well-annotated corpora. While such resources are abundant for English (e.g., RST-DT, PDTB), Chinese discourse corpora are scarce. This lack hinders research in areas like discourse segmentation and relation recognition for Chinese. Furthermore, existing Chinese corpora often lack the detailed annotation needed for advanced studies. The paper highlights that Chinese sentence structure, with its frequent use of segment separators over sentence terminators, leads to longer sentences with complex internal relationships, making annotation challenging.

Methods

The authors propose a flexible annotation framework and a web-based system to address these challenges. The system allows annotators to:

  • Select discourse units: Annotators can identify discourse pairs, which can range from clauses to multiple sentences.
  • Annotate discourse relations: They utilize the three-level hierarchical discourse relation tags from the PDTB 2.0, capturing both explicit and implicit relations.
  • Identify discourse markers: Annotators highlight discourse markers at the character level, distinguishing explicit from implicit relations.
  • Determine argument boundaries: The system supports annotating nested discourse relations by allowing annotators to define argument boundaries for complex structures.
  • Annotate sentiment polarity: Crucially, the system also captures sentiment polarity (positive, neutral, negative) for each argument and the entire discourse pair. This is done at a pragmatic level, considering the actual meaning rather than surface semantics.

The web-based interface is designed for efficiency and accuracy, allowing for repeated selection of clauses to handle nested relations.

Datasets

The paper describes the development of two pilot Chinese discourse corpora based on the Sinica Treebank 3.1.

  • First Corpus: Annotated at the inter-sentential level, focusing on the top four PDTB classes (Temporal, Contingency, Comparison, Expansion) between successive sentences.
  • Second Corpus: Annotated at the intra-sentential level, focusing on clauses and specifically on Contingency and Comparison relations within sentences.

The corpus developed using the proposed system is a new benchmark that aims for finer-grained annotation, including sentiment analysis, and addresses limitations of previous efforts.

Key Insights and Ideas

  • Chinese Sentence Structure Complexity: The high ratio of segment separators to sentence terminators in Chinese leads to longer sentences with more complex internal relationships, posing a significant challenge for discourse annotation.
  • Need for Flexible Annotation: A rigid annotation scheme is insufficient for Chinese. The proposed framework's flexibility in defining discourse units and handling nested relations is crucial.
  • Discourse Markers in Chinese: Identifying Chinese discourse markers is non-trivial due to their diverse grammatical forms and the potential for general function words to be used as markers. Automatic detection is challenging, necessitating manual annotation at the character level.
  • Sentiment-Discourse Correlation: Annotating sentiment polarity alongside discourse relations opens up new avenues for research into how sentiment is conveyed and transitions within discourse. This is a novel aspect not typically found in standard discourse corpora.
  • Pragmatic Sentiment Annotation: The emphasis on judging sentiment polarity at a pragmatic level (real meaning) rather than surface semantics is key to capturing the true sentiment conveyed in discourse.

Key Results

The paper focuses on the development of the annotation system and framework rather than presenting quantitative results from a large-scale corpus. However, it demonstrates the system's capability to:

  • Capture explicit and implicit discourse relations at various linguistic levels.
  • Handle nested discourse structures.
  • Annotate discourse markers at the character level.
  • Record sentiment polarities for discourse pairs and their arguments.

The system is presented as a tool to facilitate fast and accurate annotation, thereby reducing the cost and time associated with creating detailed discourse corpora.

Limitations and Future Work

  • Annotation Cost: The authors acknowledge that detailed annotation is time-consuming and costly, even with the developed system.
  • Discourse Marker Ambiguity: The paper notes the difficulty in automatically identifying Chinese discourse markers and the need for disambiguation.
  • Future Work: The paper implicitly suggests that the developed corpus and system will be valuable for future research in Chinese discourse processing and sentiment analysis, particularly in exploring the correlations between discourse relations and sentiment. The system's flexibility also implies potential for adaptation to other annotation tasks.

Summary updated

BibTeX
@inproceedings{huang2012annotation,
  author = {Hen-Hsen Huang and Hsin-Hsi Chen},
  title = {{An Annotation System for Development of Chinese Discourse Corpus}},
  booktitle = {International Conference on Computational Linguistics},
  year = {2012},
}
Figure