Cross-Domain Semantic Similarity

• Cross-domain semantic similarity is defined as a function measuring latent semantic alignment between data samples from different domains, independent of surface features.
• Embedding- and graph-based methods, clustering, and contrastive learning effectively assess semantic convergence across heterogeneous modalities.
• Applications range from transfer learning and domain adaptation to federated recommendations, enabling improved cross-modal translation and benchmarking.

Cross-domain semantic similarity quantifies and exploits the extent to which samples, features, or structures share meaning across different data domains. This concept is pivotal for transfer learning, domain adaptation, cross-modal translation, benchmarking evaluation, and knowledge transfer in heterogeneous or federated settings. Unlike intra-domain similarity, which presupposes shared data statistics and vocabularies, cross-domain semantic similarity must account for both distributional divergence and deeper semantic equivalence independent of surface forms, modalities, or lexical overlaps.

1. Formalizations and Notions of Cross-Domain Semantic Similarity

Definitionally, cross-domain semantic similarity is instantiated as a function $\mathsf{sim}$ that, given two objects $x \in \mathcal{X}_s$, $y \in \mathcal{X}_t$ drawn from disparate domains, returns a real-valued score or a binary equivalence label: $$\mathsf{sim}: \mathcal{X}_s \times \mathcal{X}_t \rightarrow \mathbb{R}.$$ The similarity measure may be operationalized via cosine similarity in embedding spaces, binary equivalence assertions, or graph-driven criteria. Crucially, it attempts to capture latent alignment of meaning even under significant domain shifts in data distribution, feature support, or modality.

In knowledge-driven settings, semantic similarity is assessed by mapping both domains to a semantic backbone such as a knowledge graph, facilitating comparison at the level of relations and not merely features (Wei et al., 25 Nov 2025). In representation learning, shared or aligned embeddings ground semantic similarity through explicit losses and constraints, ensuring that semantically related samples are proximal in the learned space regardless of origin (Guo et al., 20 Apr 2026, Yu et al., 2022, Ye et al., 23 Oct 2025, Bollegala et al., 2015).

2. Methodological Approaches for Cross-Domain Semantic Similarity

Methodologies for modeling and leveraging cross-domain semantic similarity can be organized by their underlying mechanisms and application scope:

• Embedding-Based Methods: Both classic (word2vec, GloVe) and modern contextual models (ELMo, USE, transformer-based sentence encoders) generate representations per domain. Similarity is typically computed as cosine or angular distance between embeddings, often averaged across a set of pivot features (e.g., adjectives, sentiment-laden tokens) (Sheoran et al., 2020, Bollegala et al., 2015). Alignment is enforced either by constraints or domain adaptation objectives (Ye et al., 23 Oct 2025, Guo et al., 20 Apr 2026).
• Clustering and Prototyping: Prototypes—vectors summarizing class or cluster semantics—are learned per domain, with inter-domain similarity promoted by contrastive loss or prototype-pair regularization. Dictionaries maintaining class-wise prototypes from multiple domains enable explicit positive-pairing of like classes and repulsion of unlike classes in the embedding space (Ye et al., 23 Oct 2025).
• Cross-Layer and Cross-Modal Alignment: In adaptation scenarios like vision or translation, semantic similarity is enforced at multiple granularities (region, class, layer) via cross-similarity maps (e.g., region-level cosine similarity using pretrained ViT tokens (Yu et al., 2022)) or prototype alignments, often accompanied by dynamic attention mechanisms or multiscale discriminators (Ma et al., 2022, Chen et al., 2023).
• Contrastive and Consistency Learning: Global and local semantic similarity is encouraged through bidirectional InfoNCE losses, aligning representations from global (pretrained) and local (client/domain-specific) views (Guo et al., 20 Apr 2026). Consistency losses further regularize models to produce comparable outputs for semantically similar but distributionally divergent inputs.
• Utility-Driven Semantic Alignment: In reasoning-intensive settings, semantic similarity is redefined as the utility of a demonstration to an LLM inference process, measured by the improvement in error over zero-shot baselines. Deltas of such utility feed into an explicit alignment loss (Ye et al., 10 Dec 2025).
• Benchmark Construction via Knowledge Graphs: Systematic construction of cross-domain semantic similarity evaluation involves generating and perturbing subgraphs, mapping to language with LLMs, and verifying equivalence by reconstructing the semantic backbone (Wei et al., 25 Nov 2025).

3. Quantification and Evaluation Metrics

A rigorous evaluation regime involves both continuous and binary measures. The most common metric is cosine similarity, particularly in embedding-based approaches, formalized as: $$\cos(\mathbf{u}, \mathbf{v}) = \frac{\mathbf{u} \cdot \mathbf{v}}{\|\mathbf{u}\| \|\mathbf{v}\|}.$$ Thresholding for equivalence yields binary predictions, scored with $F_1$, precision, and recall (Wei et al., 25 Nov 2025). Rank correlation ($\rho$) assesses continuous similarity relative to ground-truth or human evaluation. In clustering-based recommendations, metrics such as Recall@K, NDCG@K, and normalized ranking accuracy are used to relate domain similarity to downstream transfer performance (Sheoran et al., 2020, Guo et al., 20 Apr 2026).

Benchmark datasets such as Semantic-KG assess how well similarity metrics generalize across domains (general, biomedical, finance, biology), perturbation types (node/edge removal/replacement), and representation models, revealing nuanced trends and failure modes (Wei et al., 25 Nov 2025).

4. Applications in Vision, Language, and Recommender Systems

• Unsupervised Domain Adaptation and Cross-Modality Segmentation: Class-prototype alignment, similarity-based constraints, and dictionary-driven contrastive losses markedly improve adaptation across imaging domains, e.g., MRI↔CT segmentation (Ye et al., 23 Oct 2025).
• Region Similarity Matching for Video Translation: Cross-domain region similarity matching using ViT token embeddings underpins improved unpaired nighttime infrared→daytime visible video translation, with downstream benefits in detection and fusion (Yu et al., 2022).
• Federated Recommendation: Privacy-preserving cross-domain knowledge transfer is achieved by aligning textual semantics across item corpora without user overlap, using server-side semantic clustering, local adaptation, and contrastive graph learning (Guo et al., 20 Apr 2026).
• Cross-Domain Sentiment Analysis: Embedding-based and statistical similarity metrics enable informed source domain selection, improving transfer performance for sentiment classification across diverse domains (Sheoran et al., 2020).
• LLM Benchmarking and Log Anomaly Detection: Knowledge graph-driven benchmarks rigorously test the generalizability of semantic similarity measures across application domains, while utility-aligned representation spaces facilitate semantic transfer in challenging log anomaly settings (Wei et al., 25 Nov 2025, Ye et al., 10 Dec 2025).

5. Empirical Comparisons, Limitations, and Observed Trends

Multiple studies establish that cross-domain semantic similarity is highly context- and task-sensitive. No single method universally dominates. Transformer-based sentence embeddings (e.g., USE, ELMo) offer high cross-domain recall but require significant computational resources (Sheoran et al., 2020). Traditional metrics, such as BERTScore and ROUGE, can outperform LLM-as-judge in entity-level perturbations, while LLMs excel in subtle relational shifts (Wei et al., 25 Nov 2025).

In federated and UDA settings, methods that combine global semantic clustering, local fine-tuning, and graph-based fusion consistently achieve higher recall and NDCG than approaches lacking multi-level semantic alignment (Guo et al., 20 Apr 2026). Dictionary-based prototype contrast and explicit hybrid pseudo-labeling significantly boost cross-modality transfer for vision tasks (Ye et al., 23 Oct 2025, Chen et al., 2023).

Empirical studies also highlight the risk of domain overfitting: models aligned on general knowledge may degrade on domain-specific KGs or vocabularies. Systematic ablations confirm that explicit semantic similarity constraints, rather than mere distributional matching, are essential for robust cross-domain performance (Ye et al., 10 Dec 2025, Wei et al., 25 Nov 2025).

6. Open Challenges and Future Directions

Key limitations include the dependence on high-quality semantic backbones (e.g., KGs, pretrained models), cost scalability for LLM-driven validation (Wei et al., 25 Nov 2025), and the sometimes limited granularity of similarity measures (most methods focus on entity or relation levels, with less coverage of pragmatic facets like tone or context).

Future work is moving toward:

• Adversarial and richer perturbations for benchmark generation,
• Graded similarity scales (e.g., graph distance to quantify semantic drift),
• Domain adaptation for more specialized scientific domains,
• Integration of cross-modal and cross-layer attention mechanisms for deeper semantic calibration (Ma et al., 2022).

The field remains dynamic, with new architectures and analytical tools continually re-defining what it means to capture and deploy semantic similarity across heterogeneous domains in learning systems.