Identifying communicative functions in discourse with content types

Area of Science:

• Computational Linguistics
• Discourse Analysis
• Natural Language Processing

Background:

• Texts are complex structures composed of smaller communicative units (micro illocutionary acts).
• Identifying these units is crucial for understanding text purpose and extracting relevant information.
• Existing methods often lack the fine-grained analysis needed for clause-level text segmentation.

Purpose of the Study:

• To propose an empirical framework for modeling micro illocutionary acts at the clause level, termed "content types."
• To develop and validate a computational approach for automatically classifying these content types.
• To create a new, richly annotated corpus for training and evaluating the proposed models.

Main Methods:

• Development of a dedicated annotation scheme based on linguistic theories of text types (Werlich, 1976).
• Creation of a new corpus of 279 documents (over 180,000 tokens) annotated for content types.
• Implementation of a bi-LSTM model for automatic classification of content types, achieving high inter-annotator agreement (Cohen's kappa = 0.89).

Main Results:

• The proposed framework achieved an average F1 score of 74.99% for automatic content type classification.
• The model demonstrated consistent performance across contemporary and historical documents.
• Genre-specific performance varied, indicating areas for future refinement.

Conclusions:

• The study presents a robust computational framework for discourse-oriented information extraction.
• The newly annotated corpus and validated model facilitate cross-disciplinary research through computationally-aided linguistic analysis.
• This work advances the understanding of text structure and communicative functions at a micro-level.