When we process text computationally, we naturally gravitate toward the sentence as our fundamental atomic unit. Sentences have clear boundaries, explicit punctuation, and are readily handled by standard tokenizers. However, treating a sentence as an indivisible unit of information introduces a fundamental bottleneck in natural language processing: sentences are often too coarse.
Consider a complex sentence: "Although the battery life of this laptop is exceptional, the screen is dim and the keyboard feels cheap." If a model operates strictly at the sentence level, it must assign a single sentiment label, a single topic, or a single logical node to this entire string. The exceptional battery life gets averaged out or conflated with the poor screen and keyboard.
To solve this, a rich body of research focuses on sub-sentential segmentation—breaking sentences down into their smallest functional propositions, most notably Elementary Discourse Units (EDUs), independent clauses, or syntactic phrases.
What is an Elementary Discourse Unit?
An EDU is a clause-like fragment of text that represents a single, coherent thought, action, or proposition. Unlike arbitrary word n-grams or tokens, EDUs preserve semantic and functional integrity. They serve as the foundational building blocks of discourse theories, such as Rhetorical Structure Theory (RST).
By shifting the unit of analysis from the sentence to the EDU or clause, models gain a surgical precision that fundamentally changes how text is summarized, analyzed, and reasoned over.
Simultaneous Extraction and Compression in Summarization
Traditional extractive summarization models select the "top-$k$" sentences from a document. The problem with this approach is that sentences often contain redundant modifiers, parenthetical remarks, or conversational noise alongside their core facts.
By shifting the unit of selection to EDUs or sub-sentential clauses, models can perform content selection and compression simultaneously. As shown by Ching-Hsun Tseng in EDU-level Extractive Summarization with Varying Summary Lengths, extracting information at the EDU level significantly increases the theoretical upper bound of summary quality. It allows the model to retain essential clauses while cleanly discarding uninformative or redundant phrases.
Similarly, Zhenwen Li and Wenhao Wu demonstrate in Composing Elementary Discourse Units in Abstractive Summarization that using clauses as the primary building blocks allows abstractive models to exclude trivial details and seamlessly fuse information from multiple source sentences into a single, high-density output. This avoids the information loss associated with "blindly" compressing full sentences after selection, a challenge originally highlighted by Stephen Wan and Cécile Paris in Experimenting with Clause Segmentation for Text Summarization.
This fine-grained filtering is also highly effective for managing input length constraints in modern architectures. For instance, Shiyin Tan et al. utilize EDUs as latent queries in A Unified Retrieval Framework with Document Ranking and EDU Filtering for Multi-document Summarization to surgically remove irrelevant clauses while preserving the core informational signal.
Disentangling Mixed Sentiment and Bias
In sentiment analysis and bias detection, sentence-level aggregation frequently obscures the truth. When a user writes a review or a journalist writes an article, they often weave multiple perspectives or conflicting points into a single sentence.
In Multiple Instance Learning Networks for Fine-Grained Sentiment Analysis, Mirella Lapata and Stefanos Angelidis show that decomposing text into EDUs allows models to isolate specific positive or negative claims that would otherwise be averaged out within a complex sentence. This granularity is echoed by Jingjing Wang et al. in Human-Like Decision Making: Document-level Aspect Sentiment Classification via Hierarchical Reinforcement Learning, where clause-level segmentation prevents the conflation of distinct opinions regarding different aspects of a subject.
Beyond sentiment, this precision is vital for uncovering ideological framing. Nicolas Devatine et al. apply this in Predicting Political Orientation in News with Latent Discourse Structure to Improve Bias Understanding, demonstrating that segmenting text into EDUs captures subtle argumentative nuances and indicators of media bias that lexical or sentence-level models miss entirely.
Mapping the Logical Skeleton of Text
Why do sub-sentential units work so well for complex reasoning? Because arguments are built on relationships between propositions, and these relationships frequently happen inside sentence boundaries.
When a model treats an entire sentence as an atomic node, it misses internal logical cues. In DAGN: Discourse-Aware Graph Network for Logical Reasoning, Yinya Huang et al. argue that sub-sentential clauses represent the optimal level of granularity for logical reasoning graphs. By treating EDUs as nodes and discourse relations (like contrast or causation) as edges, models can map out the hidden logical skeleton of an argument.
This is particularly powerful for parsing complex rules or administrative texts. Yifan Gao et al. demonstrate in Discern: Discourse-Aware Entailment Reasoning Network for Conversational Machine Reading that breaking down rule texts into clause-like EDUs allows a system to evaluate multiple distinct logical conditions embedded within a single sentence, leading to highly precise conversational reasoning. Similarly, Xin Tan et al. leverage EDUs in Towards Discourse-Aware Document-Level Neural Machine Translation to map logical relationships via structural dependency graphs, significantly improving document-level translation coherence.
Alternative Granularities: Phrases and Tuples
While EDUs and clauses represent propositional units, some tasks benefit from shifting the granularity even further down to syntactic phrases or semantic tuples.
For example, when generating presentation slides from academic papers, full sentences create overly dense text blocks. Sida Wang et al. show in Phrase-Based Presentation Slides Generation for Academic Papers that extracting noun and verb phrases via a syntactic parser yields concise bullet points that closely match human authoring styles. In a similar vein, Lidong Bing et al. use phrase-level selection and merging in Abstractive Multi-Document Summarization via Phrase Selection and Merging to isolate specific facts and recombine them across document boundaries.
Alternatively, the authors of Extractive Summarization via Fine-grained Semantic Tuple Extraction bypass traditional sentence boundaries entirely by extracting predicate-argument tuples, allowing models to focus purely on core information structures.
Conclusion
Whether utilizing EDUs, independent clauses, or syntactic phrases, the underlying intuition remains the same: the sentence is a human grammatical convention, not a clean semantic atom. By decomposing text into sub-sentential units, NLP models can prune noise, isolate conflicting sentiments, and explicitly map the logical transitions that form the foundation of complex discourse.