Imagine trying to highlight a specific recipe ingredient in a cookbook, but your highlighter can only color entire paragraphs. You would successfully save the ingredient, but you would also be forced to preserve the backstory of the chef's grandmother.
In natural language processing, a similar bottleneck has long plagued tasks like text summarization, question answering, and information retrieval. Historically, the traditional "atomic" unit for text extraction and alignment has been the full sentence. However, sentences are rarely monolithic; they are complex structures that aggregate multiple distinct facts, parenthetical remarks, and linguistic fluff.
When a model operates strictly at the sentence level, it faces an inherent inefficiency: to extract a single salient fact, it must ingest the entire sentence container. This coarse granularity introduces irrelevant noise and rapidly consumes strict word or token budgets. To solve this, a prominent research direction shifts the focus from full sentences to sub-sentential units—such as clauses, propositions, and phrases—enabling models to isolate and align high-value information with pinpoint precision.
The Sentence-Level Granularity Bottleneck
The limitations of sentence-level extraction manifest clearly when systems operate under strict length constraints. As Stephen Wan et al. observed in Experimenting with Clause Segmentation for Text Summarization, forcing a summarizer to select full sentences inevitably packs the output with irrelevant filler text, leaving less space for actual target information.
This lack of precision creates distinct problems across different language tasks:
- Question Generation: In reading comprehension, Xinya Du et al. pointed out in Identifying Where to Focus in Reading Comprehension for Neural Question Generation that sentence-level selection prevents models from focusing on specific, answer-worthy spans, resulting in overly complex or poorly targeted questions.
- Dialogue and Meetings: In conversational settings, a single turn or utterance often bundles a key decision with conversational fluff. Claire Cardie et al. demonstrated in Unsupervised Topic Modeling Approaches to Decision Summarization in Spoken Meetings that word-level or sub-utterance filtering is required to eliminate this conversational noise. Similarly, Hyun Kim et al. showed in ExplainMeetSum: A Dataset for Explainable Meeting Summarization Aligned with Human Intent that evaluating or extracting information at the level of full conversational turns is too imprecise to capture true human intent.
- Knowledge Grounding: When generating dialogue based on external knowledge, Xueliang Zhao et al. argued in There Is No Standard Answer: Knowledge-Grounded Dialogue Generation with Adversarial Activated Multi-Reference Learning that selecting entire sentences as knowledge bases introduces peripheral details that distract the model, necessitating span-level selection instead.
Deconstructing the Sentence
To overcome these constraints, researchers decompose sentences into smaller, semantically or syntactically coherent building blocks before performing any selection or alignment.
[ The weather was clear, ] [ which allowed the rescue team to land, ] [ but the winds remained high. ]
└───────── EDU 1 ────────┘ └─────────────── EDU 2 ────────────────┘ └─────────── EDU 3 ───────────┘
The literature primarily leverages three types of sub-sentential representations:
1. Elementary Discourse Units (EDUs) and Clauses
EDUs are clause-like segments that serve as the fundamental building blocks of discourse structure. By treating EDUs as the primary unit of extraction, models can prune away non-essential clauses (like parenthetical remarks or modifiers) while preserving the core message.
For example, Jiacheng Xu et al. introduced DISCOBERT in Discourse-Aware Neural Extractive Text Summarization to extract EDUs rather than full sentences, maximizing information density. In abstractive frameworks, Zhenwen Li et al. developed EDUSum in Composing Elementary Discourse Units in Abstractive Summarization, demonstrating that selecting and fusing EDUs yields less redundant summaries. To ensure these models remain factually accurate, Narjes Delpisheh et al. utilized EDUs as precise semantic guidance in Improving Faithfulness in Abstractive Text Summarization with EDUs Using BART (Student Abstract), while Ching-Hsun Tseng et al. leveraged them in EDU-level Extractive Summarization with Varying Summary Lengths to dynamically scale summary lengths without carrying along sentence-level noise.
2. Propositions
A proposition represents an atomic factual claim. In multi-document summarization, different documents often contain sentences that overlap on some facts but diverge on others.
To resolve this alignment challenge, Ori Ernst et al. developed ProCluster in Proposition-Level Clustering for Multi-Document Summarization. Instead of clustering full sentences, ProCluster extracts and clusters individual propositions. This prevents the cross-document alignment noise that occurs when a single sentence contains multiple distinct facts, allowing the model to cleanly isolate and fuse shared information.
3. Phrases, Subtrees, and Semantic Tuples
Other approaches rely on syntactic or structural decompositions. Lidong Bing et al. proposed in Abstractive Multi-Document Summarization via Phrase Selection and Merging an optimization framework that treats noun and verb phrases as the core building blocks of a summary. Taking a more structural approach, Hajime Morita et al. decomposed sentences into dependency subtrees in Subtree Extractive Summarization via Submodular Maximization. More recently, Yubin Ge et al. utilized Information Extraction (IE) techniques in Extractive Summarization via Fine-grained Semantic Tuple Extraction to identify fine-grained semantic tuples, driving sentence selection based on the density of isolated facts rather than coarse sentence-level heuristics.
Joint Selection and Compression
Why does shifting the optimization objective to sub-sentential units work so effectively? The core intuition is that it transforms text processing from a two-stage pipeline ("select a sentence, then compress it") into a joint optimization problem, where content selection and compression happen simultaneously.
In early work, Kristian Woodsend et al. demonstrated in Automatic Generation of Story Highlights that by representing documents as a collection of phrases, Integer Linear Programming (ILP) can be used to globally optimize for both salience and grammatical correctness. This allows the system to discard linguistic fluff during the selection process. This joint mechanism was extended to cross-lingual settings by Jin-Ge Yao et al. in Phrase-based Compressive Cross-Language Summarization, and applied to query-focused mobile summarization by Hiroya Takamura et al. in TTOKU Summarization Based Systems at NTCIR-10 1CLICK-2 task to maximize information density within tight screen constraints.
However, slicing sentences apart introduces a major risk: the resulting text might become ungrammatical or lose referential clarity (e.g., extracting a phrase containing an unresolved pronoun like "he"). To maintain readability, Greg Durrett et al. introduced a global optimization framework in Learning-Based Single-Document Summarization with Compression and Anaphoricity Constraints that pairs sentence compression with coreference resolution. By enforcing explicit anaphoricity constraints, their model ensures that if a sub-sentential unit containing a pronoun is selected, the unit containing its corresponding antecedent must also be included, preserving both conciseness and coherence.
Similarly, to avoid extracting overly narrow or niche facts when operating at this granular level, Asli Celikyilmaz et al. demonstrated in Discovery of Topically Coherent Sentences for Extractive Summarization that modeling hierarchical topic correlations helps sub-sentential systems maintain a balance, ensuring that the extracted fragments still reflect broad, topically coherent themes.
Modern Applications: RAG and Localized Attribution
While early sub-sentential research focused heavily on traditional extractive summarization, the exact same intuition is proving vital for modern Large Language Model (LLM) architectures, where token budgets, context windows, and factual verification act as the primary constraints.
Prompt Compression and RAG
In Retrieval-Augmented Generation (RAG), standard systems retrieve entire documents or sentences to provide context to an LLM. However, long sentences often act as "distractors" if only a single clause is relevant.
Taeho Hwang et al. addressed this in EXIT: Context-Aware Extractive Compression for Enhancing Retrieval-Augmented Generation, noting that sentence-level extraction frequently fails to remove noise when individual sentences are excessively long or contain multiple clauses with varying relevance. Conversely, Barys Liskavets et al. showed in Prompt Compression with Context-Aware Sentence Encoding for Fast and Improved LLM Inference that while token-level compression achieves the absolute minimum token count, it can destroy readability. Finding the right sub-sentential middle ground allows models to prune unnecessary adjectives and filler words while maintaining grammatical coherence for faster LLM inference.
Pinpoint Attribution
Sub-sentential granularity is also transforming how we verify whether an LLM's output is faithful to its source material. Traditional citation systems map an entire source sentence to a generated sentence. But if a source sentence aggregates information from multiple disparate sources, a sentence-level citation forces the user to parse significant amounts of irrelevant text to verify a single fact.
Sentence-Level Citation (Coarse):
Source: "Company A acquired Company B in 2021, which led to a 20% stock surge [1]."
(User must read the whole sentence to verify just the acquisition year)
Sub-Sentential Alignment (Fine-Grained):
Source: "[Company A acquired Company B in 2021] ─── Proposition 1 ─── [Citation 1]"
"[which led to a 20% stock surge] ─── Proposition 2 ─── [Citation 2]"
To provide cleaner verification, Michael Lepioshkin et al. introduced SuperPAL in Summary-Source Proposition-level Alignment: Task, Datasets and Supervised Baseline, shifting summary-to-source alignment from coarse sentence matching to fine-grained proposition matching.
Building on this concept for real-time user interaction, the authors of LAQuer: Localized Attribution Queries in Content-grounded Generation developed a framework that allows users to highlight a specific claim within AI-generated text and receive a "pinpoint" sub-sentential citation from the source document. By moving past coarse sentence-level boundaries, this approach reduces the amount of text a user must read to verify a fact by over 50%, drastically lowering the cognitive load required to fact-check model outputs.
Key Foundational Work
Automatic Generation of Story Highlights (2010) — contributed 1 concept elements to this direction.
Experimenting with Clause Segmentation for Text Summarization (2008) — contributed 1 concept elements to this direction.