Sentence-Level Style Change Detection

• Sentence-level style change detection is a method that identifies stylistic shifts at sentence boundaries using binary labeling approaches.
• It leverages techniques such as sequence labeling, contextual encoding, and boundary-focused truncation to discern subtle stylistic variations.
• Applicable to authorship analysis, discourse segmentation, and AI-generated text detection, it distinguishes style shifts from topical changes.

Sentence-level style change detection is the problem of locating the boundaries in a sequence of sentences where writing style shifts, typically as an indicator of author change or of a transition between distinct stylistic regimes. In the PAN 2025 formulation, a document is represented as $x=(s_1,s_2,\ldots,s_n)$ and mapped to a binary boundary sequence $y=(y_1,y_2,\ldots,y_{n-1})$, where $y_i\in\{0,1\}$ indicates whether there is a style change between sentence $s_i$ and $s_{i+1}$ (Schmidt et al., 1 Aug 2025). The task occupies a fine-grained position within computational authorship analysis and intersects with segmentation, discourse modeling, speaker change detection, AI-generated text provenance, framing shifts, and diachronic syntactic variation, all of which ask whether adjacent or individual sentences belong to the same source, frame, or temporal style regime (Meng et al., 2017, Wang et al., 2023, Merrill et al., 2019).

1. Task scope and granularity

Sentence-level style change detection is defined over adjacent sentence boundaries rather than over entire documents or paragraphs. In the PAN 2025 setting, each problem is a list of sentences paired with a corresponding binary array for adjacent sentence boundaries, and the data consists of continuous discussions in Reddit threads (Römisch et al., 1 Aug 2025). The appendix of the zero-shot LLM study reports that documents often contain 2–4 authors, which typically means 1–3 style changes, making the task simultaneously a boundary detection problem and a constrained segmentation problem (Römisch et al., 1 Aug 2025).

A central issue is granularity. Earlier shared-task settings worked at paragraph level: PAN 2023 labels each adjacent paragraph pair as $1$ for style change or $0$ for no style change, and the paper explicitly notes that there is no sentence-level annotation requirement in that task (Kucukkaya et al., 2023). By contrast, PAN 2025 moves the boundary unit down to individual sentences, which makes short local decisions unavoidable and exposes the detector to “stylistically shallow” units, duplicate sentences, and punctuation-only entries (Schmidt et al., 1 Aug 2025).

The datasets are also stratified by thematic difficulty. In PAN 2025, EASY documents always span multiple topics, MEDIUM documents have limited thematic diversity, and HARD documents have no thematic diversity, so the model must rely almost entirely on subtle stylistic cues rather than topic changes (Schmidt et al., 1 Aug 2025). The same easy/medium/hard logic appears in paragraph-level PAN 2023, where the hard setup likewise suppresses topical variation and forces reliance on writing style rather than content (Kucukkaya et al., 2023). This makes sentence-level style change detection not merely a finer segmentation task, but a more stringent test of whether a model can separate style from semantics.

2. Formal problem formulations

The most direct formulation is sequence labeling over sentence boundaries. The Sequential Sentence Pair Classifier (SSPC) paper expresses the task as

$$x = (s_1, s_2, \ldots, s_n) \rightarrow y = (y_1, y_2, \ldots, y_{n-1}),$$

with

$y_i \in \{0,1\},$

where each $y_i$ indicates whether there is a style change between sentence $y=(y_1,y_2,\ldots,y_{n-1})$0 and $y=(y_1,y_2,\ldots,y_{n-1})$1 (Schmidt et al., 1 Aug 2025). This formulation preserves document order and makes boundary prediction the primary object of inference.

A second formulation reduces the task to classification over adjacent units. The PAN 2023 paragraph-level system converts multi-author writing style detection into a binary natural language inference problem over consecutive paragraph pairs, using inputs of the form

$y=(y_1,y_2,\ldots,y_{n-1})$2

and predicting whether the two paragraphs were written by the same author or not (Kucukkaya et al., 2023). The text-based speaker change detection literature uses an analogous “decision point” view: given the utterance immediately before and immediately after a decision point, together with surrounding left and right context, the model predicts whether the speaker changes (Meng et al., 2017). Although developed for speakers rather than authors, this is structurally a sentence-boundary change detector.

A third formulation separates existence detection from localization. The PAN@CLEF 2018 style breach system first predicts whether a document contains any style change at all, then recursively splits the document into two halves with equal numbers of sentences and re-applies the detector to each fragment (Kopev et al., 2019). If both halves are negative, the split point is taken as the change location; recursion stops when the fragment has fewer than 20 sentences, at which point the middle point is returned (Kopev et al., 2019). This yields sentence-boundary localization without training a dedicated boundary classifier.

Taken together, these formulations show that sentence-level style change detection can be operationalized as sequence labeling, local pairwise classification, or recursive search. The differences are not merely architectural: they encode different assumptions about whether evidence is concentrated at the boundary itself, distributed across context, or only detectable at coarser document scale.

3. Representations and contextual signals

Direct sentence-level detectors emphasize contextualized sentence representations rather than isolated sentence classification. In SSPC, each sentence is encoded with a frozen pre-trained LLM, and the submitted system uses StyleDistance/styledistance as the base transformer (Schmidt et al., 1 Aug 2025). Token-level hidden states are aggregated with mean pooling,

$y=(y_1,y_2,\ldots,y_{n-1})$3

producing a sequence of sentence embeddings $y=(y_1,y_2,\ldots,y_{n-1})$4 (Schmidt et al., 1 Aug 2025). A bidirectional LSTM then contextualizes these sentence vectors into $y=(y_1,y_2,\ldots,y_{n-1})$5, adjacent contextualized vectors are concatenated,

$y=(y_1,y_2,\ldots,y_{n-1})$6

and a three-layer feedforward network with linear layers, GELU activations, and dropout predicts the boundary label with binary cross-entropy (Schmidt et al., 1 Aug 2025). The paper’s explicit rationale is that for short, repetitive, or stylistically weak sentences, context and position within the document can be more informative than the sentence in isolation.

Hierarchical encoders provide a closely related inductive bias. The text-based speaker change detector uses a two-level LSTM architecture: a sentence encoder over words and a context encoder over sentence embeddings (Meng et al., 2017). It then introduces static sentence-level attention, in which only the two critical utterances adjacent to the decision point attend to sentences on the opposite side. The final classifier consumes the concatenation

$y=(y_1,y_2,\ldots,y_{n-1})$7

where $y=(y_1,y_2,\ldots,y_{n-1})$8 and $y=(y_1,y_2,\ldots,y_{n-1})$9 are the critical utterances and $y_i\in\{0,1\}$0 are attention summaries of opposite-side context (Meng et al., 2017). The architecture is explicitly motivated by the idea that boundary decisions should compare the critical units on either side while selectively using surrounding evidence.

Boundary-centered truncation is another recurring mechanism. In the PAN 2023 NLI approach, Transition-Focused Truncation keeps the end of the first paragraph and the beginning of the second paragraph rather than preserving the starts of paragraphs (Kucukkaya et al., 2023). For a 512-token input, this means 256 tokens from the end of paragraph 1 and 256 tokens from the beginning of paragraph 2 (Kucukkaya et al., 2023). The method reflects the assumption that stylistic continuity or discontinuity is most visible near the transition itself. The reported results show that this helps in several settings, especially easy and medium, but is not universally superior: in the hard setup, the best result came from DeBERTa with warmup but without transition-focused truncation (Kucukkaya et al., 2023).

4. Transferable signals from adjacent research areas

Several neighboring research programs do not solve sentence-level style change detection directly, but they define sentence-level signals that are reusable for it. The zero-shot long-text style transfer framework ZeroStylus extracts sentence-level stylistic templates by encoding each sentence $y_i\in\{0,1\}$1 as

$y_i\in\{0,1\}$2

clustering the embeddings with DBSCAN, and storing cluster centroids as prototype sentence templates $y_i\in\{0,1\}$3 in a repository $y_i\in\{0,1\}$4 (Wu et al., 11 May 2025). Paragraph structure is modeled through hierarchical aggregation,

$y_i\in\{0,1\}$5

and new paragraph templates are added to $y_i\in\{0,1\}$6 only when

$y_i\in\{0,1\}$7

During generation, each source sentence is matched to the best sentence template by

$y_i\in\{0,1\}$8

while the paragraph is matched to the nearest paragraph template

$y_i\in\{0,1\}$9

The paper explicitly states that it does not propose a style-change detector, a classifier for boundary detection, an explicit change-point criterion, or a sentence-to-sentence alignment algorithm for detecting style shifts (Wu et al., 11 May 2025). However, it provides sentence-template similarity, paragraph-template distance, and hierarchical consistency checks that could be repurposed for detection. This suggests a two-level anomaly view in which local sentence-template mismatch and paragraph-level inconsistency are jointly informative.

Sentence-level reframing research supplies a different perspective. “Controlled Neural Sentence-Level Reframing of News Articles” formulates reframing as a fill-in-the-blank problem over three contiguous sentences,

$s_i$0

where the generated middle sentence must express a target frame $s_i$1 while remaining coherent and topic-consistent with $s_i$2 and $s_i$3 (Chen et al., 2021). The paper uses T5-base and frame-specific generators $s_i$4 over four frames: Economic, Legality, Policy + Political, and Crime (Chen et al., 2021). Its three training strategies—framed-language pretraining, named-entity preservation, and adversarial learning—show that frame change is context-dependent and multi-objective: named entities stabilize topical grounding, while adversarial learning strengthens framing at the cost of coherence (Chen et al., 2021). The work does not propose a detector, but it operationalizes sentence-level style/frame change as a contextual transformation rather than as keyword substitution.

Sentence-level provenance detection and diachronic modeling extend the same logic to other sources of variation. SeqXGPT introduces sentence-level AI-generated text detection for documents containing both human-written sentences and LLM-modified sentences, extracts word-wise log-probability lists from GPT-2, GPT-Neo, GPT-J, and LLaMA, processes them with a five-layer 1D CNN followed by two Transformer layers, and predicts a sentence label by taking the most frequent predicted word label in the sentence (Wang et al., 2023). The diachronic POS-tagging study conditions an LSTM tagger on a year embedding $s_i$5 concatenated to every word embedding $s_i$6, that is, $s_i$7, and shows that the learned year embeddings can be used to date novel sentences (Merrill et al., 2019). In both cases, the central claim is that sentence form contains source-specific or time-specific regularities that can be modeled without requiring full-document evidence.

5. Benchmarks, metrics, and empirical results

Empirical work on sentence-level style change detection uses heterogeneous metrics because the field combines segmentation, classification, and provenance tasks (Schmidt et al., 1 Aug 2025, Römisch et al., 1 Aug 2025, Kucukkaya et al., 2023, Meng et al., 2017, Wang et al., 2023, Kopev et al., 2019). PAN sentence- and paragraph-level tasks use Macro-F1; zero-shot LLM evaluation additionally reports Hamming distance and normalized Hamming distance; style breach localization uses WindowDiff and WinPR; speaker change detection reports Accuracy, Precision, Recall, and F1; and sentence-level AI-generated text detection uses Precision, Recall, and Macro-F1.

Setting	Representative result	Metric
PAN 2025 sentence-level SSPC	official test: 0.929 / 0.815 / 0.731 on Easy / Medium / Hard; abstract also reports 0.923 / 0.828 / 0.724	Macro-F1
Zero-shot LLM vs supervised baseline	Claude-3.7-Sonnet: 0.8559 / 0.8182 / 0.6612; StyleDistance + LSTM + MLP: 0.9231 / 0.8276 / 0.7240	Macro-F1
PAN 2023 paragraph-level NLI	deberta-v3-warmup-transition: 0.987 Easy, 0.812 Medium; deberta-v3-warmup: 0.770 Hard	Macro-F1
Text-based speaker change detection	Hierarchical RNN + static attention: 89.2 accuracy, 78.4 F1	Accuracy, F1
Sentence-level AI-generated text detection	SeqXGPT: 95.3 mixed-model binary, 95.7 mixed-model multiclass, 92.8 OOD	Macro-F1
Recursive style breach localization	Stacking: 0.5719 WindowDiff, 0.3395 WinP, 0.6132 WinR, 0.3302 WinF	WindowDiff, WinPR

Several patterns recur across these results. First, direct sentence-level supervised systems are strong: SSPC outperforms the zero-shot Claude baseline on easy and hard, while remaining close on medium (Schmidt et al., 1 Aug 2025, Römisch et al., 1 Aug 2025). Second, contextual boundary modeling is consistently useful: hierarchical RNNs with static attention outperform non-attention and non-hierarchical variants in speaker change detection, and DeBERTa with warmup and transition-focused truncation is the strongest validation model overall in paragraph-level PAN 2023 (Meng et al., 2017, Kucukkaya et al., 2023). Third, zero-shot LLMs are already competitive baselines: Claude-3.7-Sonnet outperforms the PAN competition baselines reported in the LLM study and reaches about 0.83 on easy and medium paragraph-level PAN 2024, with 0.618 on the hard paragraph-level dataset (Römisch et al., 1 Aug 2025).

Results outside authorship segmentation are also informative. SeqXGPT shows that sentence-level provenance detection can substantially outperform sentence-adapted document detectors, reaching 97.2 and 97.6 Macro-F1 on GPT-2 and GPT-Neo particular-model binary detection, 95.3 on mixed-model binary detection, and 95.7 on mixed-model multiclass detection (Wang et al., 2023). The diachronic POS-tagger dates decade buckets with average error 12.5 and year buckets with 21.9, compared with 26.6 and 37.5 for a feedforward baseline, and its year embeddings yield a first-principal-component time correlation of $s_i$8 versus $s_i$9 for the feedforward model (Merrill et al., 2019). These are not sentence-boundary benchmarks, but they demonstrate that sentence-level stylistic or syntactic signals can be strong enough for fine-grained source or temporal inference.

6. Limitations, misconceptions, and open questions

A persistent misconception is that style change detection is mainly topic shift detection. The zero-shot LLM study explicitly probes this issue by correlating predicted switches with semantic similarity and finds a moderate to strong negative correlation on the easy split that weakens substantially on the hard split: for sentence similarity versus switch prediction, Spearman is -0.239 on easy, -0.160 on medium, and -0.117 on hard (Römisch et al., 1 Aug 2025). The same study notes that the ground truth in hard cases often includes author changes between semantically similar sentences (Römisch et al., 1 Aug 2025). Topic and style are therefore correlated in easier cases, but they are not equivalent.

A second misconception is that a sentence can usually be classified in isolation. PAN 2025 is difficult precisely because many sentences are “stylistically shallow,” over 10% of the sentences are exact duplicates, some appear thousands of times, and hundreds of entries are only punctuation marks but are still treated as separate sentences (Schmidt et al., 1 Aug 2025). This makes local evidence sparse and explains why contextual sequence models, left-right context encoders, and boundary-centered truncation recurrently improve performance. It also clarifies why paragraph-level models cannot simply be downscaled: methods built on adjacent paragraphs cannot detect sentence-level boundaries within a paragraph, and transition-focused truncation can discard useful broader context when global style information matters, as observed in the hard paragraph-level setup (Kucukkaya et al., 2023).

A third misconception is that generation-based style transfer or reframing systems are already detectors. ZeroStylus explicitly does not provide a style-change detector, a boundary classifier, an explicit change-point criterion, or detection metrics such as ROC/F1/accuracy (Wu et al., 11 May 2025). The reframing model likewise does not learn a classifier that flags sentence-level style shifts in existing text (Chen et al., 2021). Their value is methodological: they define sentence-level template matching, paragraph-level coherence constraints, contextual framing categories, and named-entity preservation signals that can be adapted to detection.

Open questions follow directly from the current limitations. The zero-shot LLM study identifies over-segmentation as the main failure mode: the number of predicted changes has the strongest positive correlation with Hamming distance across datasets (Römisch et al., 1 Aug 2025). The SSPC paper questions whether strong performance may partly reflect macrostructure rather than pure style and proposes future benchmark designs that reduce contextual or topic cues further, randomize sentence order, or otherwise isolate intrinsic stylistic signal (Schmidt et al., 1 Aug 2025). The same LLM study argues for more robust benchmark design using out-of-domain corpora such as literary texts, historical documents, and social media beyond Reddit, and for distinguishing evaluation tracks that allow external LLM use from those that do not (Römisch et al., 1 Aug 2025). These proposals indicate that the field has moved beyond proving that sentence-level detection is possible; the current challenge is determining exactly which signals are being detected and under what conditions they generalize.