Rhetorical Structure Theory (RST)

• RST is a discourse framework that decomposes texts into Elementary Discourse Units (EDUs) and differentiates between nucleus and satellite segments.
• It employs both top-down and bottom-up neural parsing methods, enhancing applications such as summarization, sentiment analysis, and argument mining.
• RST supports multilingual discourse analysis by adapting to various languages and genres, with robust evaluation metrics and advanced tree-structured models.

Rhetorical Structure Theory (RST) provides a hierarchical, tree-structured framework for analyzing the coherence and organization of natural language texts, focusing on the functional relationships between minimal text spans called elementary discourse units (EDUs). Central to RST is the formal distinction between nucleus and satellite, enabling models to represent how some text spans are more central to the author’s communicative intent, while others provide ancillary, supporting information. RST underpins much of modern computational discourse analysis, finding foundational applications in coherence assessment, summarization, sentiment analysis, text generation, and argument mining.

1. Core Concepts and Formal Framework

RST models the discourse structure of a document as a labeled, rooted, ordered tree:

• Elementary Discourse Units (EDUs): The minimal, clause-like spans serving as the leaves of the RST tree. EDUs are typically contiguous text segments, each expressing a single proposition or function (Chistova, 2024).
• Nuclearity: Binary relations in RST separate connected spans as either Nucleus (central) or Satellite (supporting), or as multinuclear (all children equally central) (Chistova, 2024).
• Rhetorical Relations: Each internal node is labeled with a rhetorical relation type (e.g., Elaboration, Contrast, Attribution, Cause), governed by the corpus or guideline in use (e.g., RST-DT, GUM, or language-specific adaptations) (Peng et al., 2022).
• Tree Structure: The composition of EDUs and internal spans forms a full binary constituency tree, where every non-leaf node encodes nuclearity plus a discourse relation between its children. Typical formalization: for a tree $T=(V,E,\ell)$, $V$ is the set of nodes, $E$ the edges, and $\ell$ a function assigning both nuclearity and relation labels to internal nodes (Li et al., 2023).

RST requires that every text span participates in exactly one rhetorical relation as either nucleus or satellite; the root span—sometimes called the Central Discourse Unit (CDU)—is the nucleus most prominent in the discourse (Liu et al., 2023). Standard inventories for relations vary across corpora, e.g., RST-DT defines 18 coarse types, with finer-grained extensions for other languages or genres (Chistova, 2024, Peng et al., 2022).

2. Discourse Segmentation and Parsing Methodologies

2.1 EDU Segmentation

Automatic RST analysis begins with splitting raw text into EDUs. Neural approaches dominated by sequence labeling or pointer networks have supplanted earlier rule-based or CRF models:

• The segmenter is typically a (bi-)GRU or Transformer-based encoder producing token-level embeddings, with segmentation decisions made via softmaxed attention over possible boundaries (Lin et al., 2019).
• State-of-the-art segmenters (e.g., ELMo-based, pointer networks) approach or surpass 95 F1 versus human agreement (98.3 F1) on RST-DT (Lin et al., 2019).

2.2 Tree Construction (Parsing)

RST discourse parsing constructs a binary tree covering the identified EDUs. Two dominant architectures prevail:

• Top-Down Parsing: Recursively splits spans, at each step choosing a split point and labeling the resulting span pair with relation/nuclearity using pointer networks and bi-affine classifiers (Liu et al., 2020, Lin et al., 2019).
• Bottom-Up Parsing: Shift–reduce algorithms combine adjacent subtrees using neural classifiers, often implemented as stack/queue operations (Liu et al., 2023, Heilman et al., 2015).
• Joint models perform segmentation and parsing together, improving robustness to error propagation, with end-to-end complexity $O(n)$ in recent neural models (Lin et al., 2019).

Relation and nuclearity classification is typically performed via neural softmax layers over learned representations of the span pairs; classifier input representations are contextual, sometimes augmented with boundary token features or sentence-level embeddings (Shi et al., 2020).

3. Applications and Modeling Advances

3.1 Summarization with RST-Aware Mechanisms

RST features have been injected into advanced generative models for abstractive summarization. The RST-LoRA architecture integrates parser-estimated relation uncertainties as soft token-level masks within LoRA’s low-rank adaptation modules, enabling parameter-efficient yet discourse-informed finetuning. RST-aware LoRA variants (especially those propagating real-valued, relation-sensitive discourse scores) consistently outperform vanilla LoRA and full-network fine-tuning—yielding new state-of-the-art on long-document summarization benchmarks (e.g., Multi-LexSum, eLife, BookSum) with less than 0.1% of parameters tuned (Liu et al., 2024).

3.2 Coherence Evaluation

Neural recursive models over RST trees (RST-Recursive) achieve competitive coherence classification accuracy with far fewer parameters than sequential sentence-based models. Silver-standard RST features, even when parser-induced, can meaningfully improve coherence detection, particularly when nuclearity and explicit relation labels are retained (Guz et al., 2020).

3.3 Sentiment Analysis and Text Generation

RST-based sentiment models reweight discourse units either by structural depth or by recursive neural composition up the RST tree, providing measurable gains over bag-of-words and even classifier baselines. The Rhetorical Recursive Neural Network (R2N2) demonstrates that contrastive vs. non-contrastive relation distinction further improves performance (Bhatia et al., 2015). In text generation, models such as RSTGen allow direct control over the discourse structure of output by conditioning on RST trees, yielding interpretable and controllable long-form outputs (Adewoyin et al., 2022).

3.4 Argument Mining

RST-inspired dependency structures over discourse units can be leveraged for end-to-end argument mining. Exposing argument parsers to multiple plausible RST variants through paraphrase-augmented training boosts robustness and accuracy in mining support/attack structures across languages (Chistova, 2024).

4. Multilingual and Cross-Lingual RST Models

While RST was originally developed for English, major efforts have constructed treebanks and adapted annotation for Persian, Mandarin, Russian, Dutch, German, Spanish, Portuguese, and Basque (Liu et al., 2020, Peng et al., 2022, Shahmohammadi et al., 2021, Peng et al., 2022, Chistova, 2024). Multilingual neural parsers leverage pretrained encoders (e.g., XLM-RoBERTa), cross-lingual fine-tuning, and even EDU-aligned translation to yield state-of-the-art cross-lingual performance—even near monolingual upper bounds in Span and Nuclearity evaluation (Liu et al., 2020, Chistova, 2024).

Bilingual and parallel corpora (e.g., GUM-English/Russian) facilitate controlled studies of cross-lingual transfer and tree shape invariance, showing that the global discourse organization modeled by RST tends to be highly transferable, with tree construction metrics plateauing with modest in-language annotation (Chistova, 2024).

5. Challenges, Error Analysis, and Theoretical Extensions

5.1 Empirical Challenges in RST Parsing

Recent error analysis identifies the main bottlenecks in RST parsing as:

• Implicit Relations: Unmarked connections are far more error-prone than explicit (marker-cued) ones.
• Long-Distance Dependencies: Inter-sentential or inter-paragraph attachments are systematically more difficult for both bottom-up and top-down parsers, overshadowing effects of lexical overlap or out-of-vocabulary rates (Liu et al., 2023).
• Genre Variability: Parsing accuracy varies by genre, motivating domain-robust models and diverse training data.
• Structural Rigidity: Attachment errors sometimes arise from the tree constraint, suggesting that allowing multiple concurrent relations (as in eRST or SDRT) may yield analyses better aligned with human intuition (Zeldes et al., 2024).

5.2 Enhancements: eRST and Beyond

Enhanced RST (eRST) generalizes classical RST by introducing secondary (non-projective or concurrent) edges and explicit annotation of discourse signals (connectives, cues, lexical chains). An eRST analysis is a labeled, primary-tree plus signal-licensed graph, addressing RST’s limitations in representing tree-breaking phenomena and facilitating fine-grained, signal-driven relation typology (Zeldes et al., 2024).

eRST’s annotation and graph model support richer downstream applications—including more granular relation extraction, explainable summarization, and integration with LLMs for text planning and evaluation.

6. Evaluation, Tools, and Task Metrics

Evaluation of RST models follows standard micro-averaged F1 for:

• Span: unlabeled span detection,
• Nuclearity: correct nucleus/satellite labels on spans,
• Relation: correct relation assignment,
• Full: all of the above jointly.

Parseval and RST-Parseval conventions dominate; in multilingual and eRST settings, secondary edge and signal anchoring metrics are additionally used (Adewoyin et al., 2022, Zeldes et al., 2024).

Annotation and visualization tools such as rstWeb and ANNIS facilitate tree and graph annotation, search, and inspection; neural parsing systems are widely open-sourced, covering most high-resource and some lower-resource languages.

7. Outlook and Research Directions

RST remains a foundational framework for discourse modeling. Contemporary research seeks to extend its scope via cross-lingual and multilingual transfer, integration into parameter-efficient neural architectures, joint modeling with argument structure, and enhancement with graph constructs (eRST).

Key open challenges include:

• Developing parsers robust to long-distance and implicit relations, potentially via global or multi-pass inference (Liu et al., 2023).
• Harmonizing relation inventories for true cross-lingual universality (Chistova, 2024).
• Incorporating non-tree (graph) structures and explainable signal-cue annotation, particularly for dialogic and multiparty discourse (Zeldes et al., 2024).
• Leveraging RST-encoded structures for controllable and interpretable text generation (Adewoyin et al., 2022).
• Bridging the gap between symbolic (tree-based) and neural (latent) representations in LLMs, including by directly prompting large-scale generative models in the style of current SOTA discourse parsers (Maekawa et al., 2024).

RST’s influence continues to expand, fundamentally shaping approaches to discourse representation, automatic summarization, argument mining, sentiment analysis, and discourse-level evaluation across languages and genres.