Traditional Natural Language Processing often treats documents as monolithic sequences of tokens. However, professional and academic texts are rarely uniform; they are composed of distinct functional segments—such as "Background," "Methods," "Facts," or "Conclusions"—each serving a specific rhetorical purpose. Functional decomposition is the process of partitioning these unstructured documents into their constituent logical units to enable more precise analysis, summarization, and information extraction.
The Grammar of Rhetorical Moves
At the heart of functional decomposition is the concept of "rhetorical moves." As Wu et al. describe, these moves represent the underlying logic of a composition, such as the transition from a "Purpose" statement to a "Results" summary in a scientific abstract. By identifying these moves, researchers can map linguistic collocations to structural stages, providing a framework for understanding how professional documents are constructed.
This structural mapping is particularly useful in education and cross-cultural linguistics. For instance, Zhou et al. applied move analysis to help international students master the conventions of academic Chinese writing. Similarly, D. Huang demonstrated that while different writer groups might follow similar structural moves, their reliance on specific "lexical chunks" varies, with some using formulaic phrases for structural coherence and others for descriptive detail.
Partitioning via Rules and Patterns
In domains with highly standardized formats, such as law and finance, decomposition often relies on rule-based segmentation. By leveraging regular expressions and keyword patterns, researchers can reliably isolate specific thematic sections from large, unstructured corpora.
- Legal Documents: In the HLDC corpus, Kapoor et al. and the authors used regular expressions to partition Indian court rulings into metadata, factual bodies, and judicial results. Similarly, W. Huang et al. employed an automated pipeline to segment Chinese legal documents into factual descriptions and statutory citations, while Y. Huang et al. used keyword patterns to isolate claims and court views in civil appeal cases.
- Financial and Technical Texts: Foroutan et al. applied regular expressions to German fund prospectuses to extract sections related to company outsourcing. In the technical domain, Symonenko et al. used sublanguage theory to develop logical rules that segment utility "trouble tickets" into actionable information fields.
Enhancing Downstream Performance
Decomposing a document is rarely the end goal; rather, it serves as a critical preprocessing step that improves the performance and explainability of downstream tasks.
Summarization and Extraction: Treating a document as a set of functional parts allows models to preserve the logical flow of the original text. Shukla et al. found that rhetorical structure-aware segmentation is essential for legal summarization, as it ensures that facts, arguments, and rulings are all represented. In procedural text, Zhong demonstrated that identifying specific segments representing decision criteria allows a system to isolate discrete logical units for conversational machine reading.
Text Similarity and Retrieval: In biomedical research, Neves et al. showed that representing documents as concatenations of extracted rhetorical elements (like "Methods" or "Outcomes") significantly outperforms treating them as monolithic blocks. This allows for more nuanced similarity tasks, such as finding alternative experimental methods by focusing specifically on the "Methods" section rather than the entire paper.
Scholarly Analysis: Large-scale datasets like unarXive 2022 (Saier et al.) explicitly label text blocks by their functional roles (e.g., proofs, listings, or specific section types). This structural annotation enables researchers to build social networks of acknowledgments (Giles et al.) or analyze the power dynamics inherent in "Instructions for Authors" (X. Liu).
Beyond Sections: Narrative and Style
Functional decomposition also extends to macro-level narrative structures. Dai et al. moved beyond topical categorization to classify news articles based on journalistic structures like the "Inverted Pyramid" or "Kabob." By decomposing articles into functional building blocks like "introductory hooks" or "chronological narrations," they showed that algorithms can identify the underlying organizational logic of a story.
This structural lens even reveals how persuasion works. El Baff et al. decomposed news editorials into introductions, bodies, and conclusions to track how stylistic choices across these segments influence readers of different political ideologies. Their work suggests that the sequence of functional segments is often as predictive of persuasion as the content itself.
Key Foundational Work
Computational Analysis of Move Structures in Academic Abstracts (2006) — contributed 1 concept elements to this direction.
Illuminating Trouble Tickets with Sublanguage Theory (2006) — contributed 1 concept elements to this direction.
Towards Building and Analyzing a Social Network of Acknowledgments in Scientific and Academic Documents (2012) — contributed 1 concept elements to this direction.