APA-RST: A Text Simplification Corpus with RST Annotations

Freya HewettACL 2023

TL;DR

This paper introduces APA-RST, a novel German-language corpus of parallel newspaper articles simplified at three distinct complexity levels. The corpus is annotated with Rhetorical Structure Theory (RST) relations, offering a unique resource for understanding how text structure changes during simplification and its impact on complexity.

What Problem Does This Paper Address?

Text simplification aims to make complex texts more accessible by reducing their complexity while preserving core information. While much research has focused on sentence-level simplification, understanding text-level simplification – how coherence, discourse structure, and overall readability are affected – is crucial for creating truly accessible content. Existing resources often lack detailed discourse-level annotations, making it difficult to study these text-level phenomena. This paper addresses this gap by providing a richly annotated corpus that bridges the divide between text simplification and discourse structure analysis, specifically for the German language.

Methods

The authors leverage the Rhetorical Structure Theory (RST) framework to annotate the corpus. RST posits that text segments (Elementary Discourse Units or EDUs) are related to each other through specific rhetorical relations (e.g., 'evidence', 'elaboration'). The paper details the annotation process, which involved five annotators using the rstWeb tool. To ensure quality, a subset of the corpus was annotated by three annotators, and inter-annotator agreement was calculated. Additionally, sentence-level alignments were created between the original and simplified texts to identify which parts of the original content contribute to the simplified versions.

Datasets

The core dataset introduced is APA-RST, a corpus of 75 parallel German-language newspaper articles. These articles are sourced from the Austria Press Agency (APA) and are available at three complexity levels: original, B1 (intermediate), and A2 (basic, akin to "Leichte Sprache"). The corpus is unique in its combination of parallel texts across complexity levels and detailed RST annotations.

Key Insights and Ideas

  • Text-level simplification is more than just sentence shortening: While simplified texts are significantly shorter, the paper highlights how discourse structures change. Simplified texts tend to have more 'elaboration' and 'e-elaboration' relations, suggesting a focus on explaining concepts more directly.
  • Loss of nuance and subjectivity in simplification: The analysis of causal relations reveals that simplified texts use fewer 'reason' relations, which are typically used to link subjective claims. This suggests that simplification might reduce the expression of subjective viewpoints.
  • Structural differences in discourse organization: 'List' relations, which encompass more EDUs in original texts, appear at a higher level in the RST tree in original texts compared to simplified ones. This indicates a more hierarchical organization of enumerations in original texts.
  • Absence of 'sameunit' relations in simplified texts: These relations, used for nested constructions in original texts, are not found in the simplified versions, suggesting a flattening of complex structural embedding.
  • Alignment reveals content selection: The sentence alignments show how specific content from original texts is selected and rephrased for simplification, offering insights into what information is deemed essential.

Key Results

  • Simplified texts (B1 and A2) are approximately one-third the length of original texts.
  • A2 texts are slightly longer than B1 texts, likely due to more detailed explanations.
  • RST analysis shows increased 'elaboration' and 'e-elaboration' in B1/A2 texts compared to original texts.
  • 'List' relations in original texts encompass an average of 8.2 EDUs, compared to approximately 4 in simplified texts.
  • Attribution relations are more frequent in standard texts, indicating a reduction in the number of external perspectives in simplified versions.
  • 'Sameunit' relations are absent in simplified texts.
  • Sentence alignment Kappa scores were high: 0.77 for original-to-B1 and 0.9 for B1-to-A2.

Limitations and Future Work

The authors acknowledge that the corpus is relatively small. They also note that the study is limited to the newspaper article text type, and other text types may exhibit different simplification patterns. Future work could involve expanding the corpus to include other text types and further exploring the relationship between discourse structure and text complexity. The authors also suggest that the corpus can be used to investigate specific questions about the interaction of text complexity and discourse structure, and to develop and evaluate text simplification models.

Summary updated

BibTeX
@inproceedings{hewett2023apa,
  author = {Freya Hewett},
  title = {{APA-RST: A Text Simplification Corpus with RST Annotations}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2023},
}
Figure