Semantic Shift: A Dynamic Linguistic Change

• Semantic shift is the change in lexical meaning over time and across contexts, involving both long-term linguistic drifts and short-term cultural shifts.
• Computational methods, including contextual embeddings and alignment techniques, measure semantic shift by tracking variations in word usage and distributional patterns.
• Understanding semantic shift is crucial for applications in historical linguistics, discourse analysis, and NLP model robustness, despite challenges in evaluation and scalability.

Semantic shift denotes change in semantic organization across contexts. In diachronic linguistics, it is the change in the lexical meaning of a word over time, rather than change in grammatical function, and in computational work it is typically operationalized through changes in contextual distributions across corpora partitioned by time (Kutuzov et al., 2018). Contemporary research extends the phenomenon beyond long-range historical drift to cultural events, viewpoints, academic domains, and multi-period trajectories, while some adjacent technical literatures use the same term for representation-space drift, language-dependent offsets, or semantic deformation inside learned embeddings rather than for lexical sense change alone (Azarbonyad et al., 2017, Kiyama et al., 16 Jan 2025, Gao et al., 22 Mar 2026).

1. Canonical linguistic sense

The canonical sense of semantic shift is lexical semantic change: a word changes meaning over time. A survey of diachronic word embeddings defines semantic shift as change in lexical meaning and recalls Bloomfield’s characterization of semantic change as “innovations which change the lexical meaning rather than the grammatical function of a form” (Kutuzov et al., 2018). Within historical semantics, recurrent patterns include narrowing, broadening, pejoration, and amelioration, while computational work has often grouped phenomena into slower linguistic drifts and shorter-horizon cultural shifts (Kutuzov et al., 2018).

This distinction matters methodologically. Linguistic drifts are typically associated with decades or centuries, whereas cultural shifts may emerge over years or months in response to social and political events. The survey literature also emphasizes that semantic shift is not reducible to raw frequency change: a word may become more or less frequent without changing meaning, and meaning may shift without large frequency effects (Kutuzov et al., 2018). A closely related misconception is that semantic shift is inherently temporal. Work on political and media discourse shows that the same word can vary across viewpoints even within a short period, so semantic instability can be perspectival as well as diachronic (Azarbonyad et al., 2017).

A second core distinction concerns level of analysis. The contextualized-embedding survey defines Semantic Shift Detection as identifying, interpreting, and assessing change in the meanings of a target word, and shows that modern systems can target either global word-level change or sense- and usage-level restructuring (Montanelli et al., 2023). This has made semantic shift a bridge topic linking diachronic linguistics, lexical semantics, distributional semantics, historical NLP, and downstream applications such as information retrieval, discourse analysis, and social monitoring.

2. Distributional and embedding-based formulations

The standard computational premise is usage-based: if a word’s contexts change, its meaning has changed. Early distributional work used frequency and co-occurrence statistics, then sparse count models such as LMI, PPMI, and LSA/SVD, and later prediction-based embeddings such as SGNS, CBOW, and GloVe (Kutuzov et al., 2018). A central technical issue is comparability across time, because independently trained embedding spaces are not directly aligned.

One classical remedy is post hoc alignment, often with orthogonal Procrustes. In the survey formulation, if $\mathbf{X}$ and $\mathbf{Y}$ are anchor-word matrices from two periods, a rotation $\mathbf{R}$ is learned by

$$\min_{\mathbf{R}: \mathbf{R}^\top \mathbf{R}=\mathbf{I}} \|\mathbf{X}\mathbf{R}-\mathbf{Y}\|_F .$$

Other strategies include second-order similarity profiles, incremental initialization from one period to the next, and jointly trained dynamic embeddings with temporal coupling (Kutuzov et al., 2018).

Once vectors are made comparable, the simplest shift score is often self-distance,

$$\text{shift}(w;t_1,t_2)=1-\cos(\mathbf{w}^{(t_1)},\mathbf{w}^{(t_2)}),$$

but the literature has repeatedly shown that global vector displacement is only one view of the problem (Kutuzov et al., 2018). The contextualized survey systematizes later work along three axes: meaning representation (form-based versus sense-based), time-awareness (time-oblivious versus time-aware), and learning modality (unsupervised versus supervised) (Montanelli et al., 2023). This framework helps explain why prototype-based methods often do well on scalar ranking tasks, while sense-based methods are more interpretable but harder to compress into a single robust score (Montanelli et al., 2023).

3. Contextualized, sense-based, and instance-level modeling

Contextualized models shifted the field from type-level comparison to token-level analysis. “Analysing Lexical Semantic Change with Contextualised Word Representations” represents each occurrence of a target word with BERT, clusters usage vectors into usage types, and measures change with Entropy Difference, Jensen–Shannon Divergence, and Average Pairwise Distance; it also introduces a human similarity dataset of 3,285 usage pairs and reports significant positive correlation between model similarities and human judgments for 10 of 16 target words (Giulianelli et al., 2020). This line of work established that occurrence-level contextual geometry can reveal synchronic polysemy and diachronic change without requiring a fixed sense inventory.

A more recent development replaces balanced matching between periods with unbalanced transport between sets of contextualized instances. For a target word with old embeddings $\{\mathbf{u}_i\}_{i=1}^m$ and modern embeddings $\{\mathbf{v}_j\}_{j=1}^n$, “Quantifying Lexical Semantic Shift via Unbalanced Optimal Transport” models alignment by

$$\min_{\mathbf T\in\mathbb R_+^{m\times n}} \sum_{i,j} T_{ij}C_{ij} + \lambda_1 D_1(\mathbf T\mathbf 1_n, \mathbf a) + \lambda_2 D_2(\mathbf T^\top \mathbf 1_m, \mathbf b),$$

with cosine transport cost

$C_{ij}=1-\cos(\mathbf u_i,\mathbf v_j).$

Its central quantity, Sense Usage Shift (SUS), interprets unmatched transport mass as signal rather than nuisance: old instances with insufficient outgoing mass correspond to decreasing or disappearing senses, and modern instances with insufficient incoming mass correspond to increasing or emerging senses. On English DWUG v3, SUS achieves instance-level Spearman correlation $0.46$, outperforming LDR at $\mathbf{Y}$0 and WiDiD at $\mathbf{Y}$1, while remaining competitive on word-level magnitude and broadening/narrowing tasks (Kishino et al., 2024).

Interpretability has also motivated graph-based approaches. “Word-Centered Semantic Graphs for Interpretable Diachronic Sense Tracking” constructs a graph for each target word and time slice by combining SGNS neighbors with masked-language-model substitutes, removes the hub word, and treats connected components in the peripheral graph as sense-related communities. In the New York Times Magazine corpus, it uses this representation to show event-driven sense replacement for trump, relative stability with over-segmentation for god, and gradual digital-communication drift for post (Kolli et al., 29 Jan 2026). This suggests that semantic shift can be read not only as vector movement but also as topological change in a word’s local semantic neighborhood.

4. Temporal trajectories, viewpoints, and domain transfer

Many shifts are not well described by a single before/after comparison. “Analyzing Continuous Semantic Shifts with Diachronic Word Similarity Matrices” represents a word by an all-pairs temporal similarity matrix

$\mathbf{Y}$2

using aligned PPMI-SVD joint embeddings across arbitrary time periods. On pseudo-data with seven temporal schemas, cosine similarity plus upper-triangular features plus hierarchical clustering plus standardization reaches $\mathbf{Y}$3 accuracy, and real-data analyses on COHA and COCA distinguish enduring transitions, temporary spikes, and relative stability (Kiyama et al., 16 Jan 2025). This reframes semantic shift as a temporal pattern rather than only an adjacent difference or change point.

Semantic shift can also be conditioned by discourse communities. “Words are Malleable” defines a viewpoint as a set of texts sharing a metadata feature and compares viewpoint-specific embedding spaces with linear mappings, neighborhood-based stability, and a combined score. Its UK parliamentary and New York Times experiments show that semantic shifts occur not only over time but across political or media viewpoints, and that the combination of mapping and neighborhood evidence performs best for both intrinsic ranking and downstream tasks such as ideology detection and contrastive viewpoint summarization (Azarbonyad et al., 2017).

A related but more manual line of work examines transfer between academic fields. “On Identifying Points of Semantic Shift Across Domains” tracks “semantic evolution points” by following citations backward from early Computer Science uses of terms such as polymorphism, semaphore, and ontology into older source domains (Choi et al., 2023). Here the shift is not estimated from embeddings but from documented borrowing and reinterpretation across disciplines.

Short-horizon scientific change has likewise been studied through interpretable axes. “Semantic coordinates analysis reveals language changes in the AI field” represents a target word relative to stable coordinate pairs using

$\mathbf{Y}$4

approximated by cosine similarities of aligned embeddings, and summarizes trajectories by linear trend. It reports, for example, that deep shifts from rigorous toward neural, and that collaboration contains less competition and more communication across 2007–2016 AI publications (Zhu et al., 2020).

5. Evaluation, significance, and practical pitfalls

Evaluation remains one of the field’s hardest problems. The survey literature repeatedly notes the scarcity of gold standards, sensitivity to corpus composition, frequency effects, and the difficulty of distinguishing genuine semantic change from genre or sampling artifacts (Kutuzov et al., 2018, Montanelli et al., 2023). Contextualized benchmarks such as SemEval-2020, DIACR-Ita, RuShiftEval, LSCDiscovery, and NorDiaChange mainly evaluate scalar ranking by Spearman correlation, which tends to favor form-based methods that naturally output one score per word (Montanelli et al., 2023).

Several papers respond by making uncertainty explicit. “Statistically significant detection of semantic shifts using contextual word embeddings” combines contextual BERT representations with permutation tests and Benjamini–Hochberg false discovery rate control. In simulation, FDR filtering returns about 177–189 words per run and reaches precision at 177 of at least $\mathbf{Y}$5, compared with $\mathbf{Y}$6 for an optimized term-frequency baseline; on downsampled SemEval settings it produces the highest or joint-highest Spearman correlation in 8 of 11 experiments (Liu et al., 2021). The implication is that semantic shift scores should be read as effect estimates with sampling variability, not as self-validating facts.

The downstream consequences of unstable semantics can be substantial. In longitudinal mental-health monitoring, “The Problem of Semantic Shift in Longitudinal Monitoring of Social Media” uses neighborhood overlap

$\mathbf{Y}$7

to rank terms by semantic stability, then shows that adding only a small number of semantically unstable features can materially alter estimated changes in depression prevalence; in one highlighted setting, the minimum and maximum Multi-Task Learning estimates differ by nearly $\mathbf{Y}$8 even though predictive performance is nearly flat across large vocabulary ranges (Harrigian et al., 2022). This makes semantic shift a robustness problem, not merely a descriptive one.

Event-driven corpora illustrate the same point on shorter horizons. “How COVID-19 Is Changing Our Language” aligns monthly Twitter embeddings with a pre-COVID reference by orthogonal Procrustes and defines a back-and-forth rotational stability score

$\mathbf{Y}$9

with round-trip cosine similarity after forward and inverse mappings. It shows that words such as racism, hero, quarantine, and ai move toward pandemic-specific neighborhoods between April and June 2020, and that the cross-corpora shift is stronger than month-to-month change within COVID discourse (Guo et al., 2021).

6. Broader technical uses of the term

Outside lexical semantics, “semantic shift” has become a broader technical label for semantically meaningful mismatch between representation spaces, tasks, or communication levels. In communications, “Towards Semantic Communications: A Paradigm Shift” uses the phrase to denote a transition from Weaver’s LEVEL-A technical fidelity to LEVEL-B semantic fidelity, where the objective becomes meaning preservation or task completion rather than exact symbol reproduction (Niu et al., 2022).

In multilingual dense retrieval, “SHIFT” treats cross-language bias as a language-specific offset in embedding space. Using parallel translation pairs, it estimates a relative language vector

$\mathbf{R}$0

and corrects indexed document embeddings by subtracting this vector. Across four MLIR benchmarks, this training-free index-side transformation improves both conventional retrieval metrics and Target-Languages Recall@20, thereby interpreting language bias as a semantically harmful shift in representation geometry (Jang et al., 17 Jun 2026).

In text embedding theory, “Semantic Shift: the Fundamental Challenge in Text Embedding and Retrieval” defines semantic shift inside a document as local semantic evolution times global semantic dispersion,

$\mathbf{R}$1

and proves that pooled text embeddings necessarily move away from constituent sentence embeddings as sentence-level diversity increases. Its controlled repeat, sequential, and random concatenation experiments argue that retrieval degradation is predicted by internal semantic shift more closely than by text length alone (Gao et al., 22 Mar 2026).

In continual learning and domain adaptation, the term marks representation drift that disrupts transfer. “Semantic Shift Estimation via Dual-Projection and Classifier Reconstruction” defines semantic shift in exemplar-free class-incremental learning as the movement of old-task embeddings after learning new tasks, estimates it with a task-wise linear map plus a category-wise row-space projection, and uses the corrected statistics for classifier reconstruction (He et al., 7 Mar 2025). “Alleviating Semantic-level Shift” uses the phrase for class-wise mismatch between source and target features in semantic segmentation and adds semantic-level adversarial alignment to complement global domain alignment (Wang et al., 2020). These uses are not about lexical meaning change, but they preserve the core intuition of semantically structured mismatch between distributions or representations.

7. Open problems

Several problems remain persistent across these literatures. The diachronic surveys call for broader language coverage, stronger gold standards, richer meaning descriptions, and more community infrastructure, noting that the field has been fragmented across NLP, IR, and neighboring disciplines (Kutuzov et al., 2018, Montanelli et al., 2023). Contextualized methods still face a trade-off between scalar performance and interpretability: form-based measures often rank words well, whereas sense-based and usage-based methods better reveal what changed but are more fragile computationally and methodologically (Montanelli et al., 2023).

Scalability is another recurring constraint. Token-level contextual methods require large numbers of embeddings, clustering or transport over many instances, and often pairwise costs. The UOT-based SUS framework is tractable on English DWUG’s roughly 100-by-100 instance regime, but the authors explicitly note that scaling to much larger corpora would require care (Kishino et al., 2024). Multi-period methods likewise expose the cost of moving beyond adjacent comparisons, even when they use lightweight embeddings (Kiyama et al., 16 Jan 2025).

A final unresolved issue is the relation between semantic shift and external validity. Some methods target lexical meaning change in the narrow historical sense; others target event-driven contextual drift, viewpoint-specific framing, or representation-space deformation in learned systems. This suggests that “semantic shift” is now an umbrella term whose precise meaning depends on what is taken to shift: word senses, sense frequencies, discourse neighborhoods, document-internal semantics, language offsets, or class embeddings. The common denominator is not a single algorithmic form, but the idea that semantically organized structure changes across time, domains, viewpoints, or learning stages, and that such change must be measured at the level where the relevant semantic organization is actually represented.