Baselines for Detection and Classification of Discourse Presentation in English Narrative

2026-06-25 · Source: Paper Index on ACL Anthology · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Natural Language Processing · Depth: Advanced, medium

Summary

Reinaldo Di Polo, Mustafa Ocal, and Mark Finlayson's 2026 paper introduces baseline computational approaches for detecting and classifying discourse presentations in English narratives. Discourse presentation involves identifying speech, writing, or thought (SW&T) attributed to a character, categorized as either direct (verbatim) or indirect (narrator's words). The research frames this as a five-way clause classification task: Direct Speech & Writing, Direct Thought, Indirect Speech & Writing, Indirect Thought, and Narrative. The authors evaluated four methods on a corrected Semino & Short's English Narrative SW&TP corpus: a CNN, Claude Sonnet 4.6, untuned BERT, and fine-tuned BERT. Fine-tuned BERT achieved the strongest overall performance with an 0.88 F1 score, significantly outperforming Claude's 0.71 F1 and CNN's 0.43 F1. However, fine-tuned BERT still showed concentrated errors in the Indirect Speech & Writing category, achieving an F1 of 0.60. The authors released their code and the corrected dataset to promote reproducibility.

Key takeaway

For NLP Engineers developing narrative analysis tools, you should prioritize fine-tuned BERT models for discourse presentation classification, given their 0.88 F1 performance. Be aware that indirect speech and writing categories remain challenging, with an F1 of 0.60, requiring focused attention during model refinement. Utilize the released code and corrected Semino & Short's English Narrative SW&TP corpus to accelerate your development and ensure reproducibility.

Key insights

Fine-tuned BERT significantly improves discourse presentation classification, though indirect forms remain challenging.

Principles

• Indirect discourse detection is complex.
• Fine-tuning models enhances performance.
• Reproducibility requires code and data.

Method

The task is framed as a five-way clause classification: Direct Speech & Writing, Direct Thought, Indirect Speech & Writing, Indirect Thought, and Narrative, evaluated on a corrected Semino & Short's English Narrative SW&TP corpus.

In practice

• Use fine-tuned BERT for discourse classification.
• Focus error analysis on indirect speech/writing.
• Utilize the released dataset for research.

Topics

• Discourse Presentation
• Narrative Analysis
• BERT
• Natural Language Processing
• Text Classification
• Semino & Short Corpus

Best for: Research Scientist, AI Scientist, NLP Engineer, Machine Learning Engineer

Related on AIssential

• Comparing BERT Sentence-Pair Classification and Few-Shot LLM Prompting for Detecting Threat and Solution Framing in German Climate News — Fine-tuned BERT models outperform few-shot LLMs for German climate news framing detection, especially with sentence context.
• A Cross-Genre Analysis of Discourse Relation Signaling in the GUM Corpus — The paper analyzes discourse relation signaling across diverse genres in the GUM Corpus.
• Few Shades of Supervision for Discourse Segmentation — Classical fine-tuning of LLMs is most effective for EDU segmentation in spontaneous speech.
• Characterizing Narrative Content in Web-scale LLM Pretraining Data — Narrative structure in web-scale LLM pretraining data is measurable and unevenly distributed, impacting model reasoning.
• Fine-Tuning BERT for Sentiment Analysis: A Beginner-Friendly Guide — Fine-tuning pre-trained transformer models like BERT enables high-accuracy NLP tasks with minimal training data.
• Enhancing Science Classroom Discourse Analysis through Joint Multi-Task Learning for Reasoning-Component Classification — Automated discourse analysis using LLM augmentation improves utterance classification and reveals effective teaching patterns.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Paper Index on ACL Anthology.