Context-Sensitive Similarity Computation

Updated 11 May 2026

Context-Sensitive Similarity Computation is a framework that integrates contextual cues—linguistic, ontological, or multimodal—into standard similarity measures to capture nuanced semantic differences.
It employs methodologies such as contextual embeddings, hybrid models, and metric learning to dynamically adapt similarity assessments based on the situation or structure.
Applications of this approach span NLP, information retrieval, computer vision, and formal reasoning, enabling facet-aware retrieval, analogical matching, and enhanced disambiguation.

Context-sensitive similarity computation encompasses a diverse set of methodologies that generalize beyond static, global notions of similarity. These approaches incorporate external or internal context—linguistic, multimodal, ontological, or referential—into the similarity assessment, enabling nuanced distinctions driven by situation, intention, or structure. Such models underpin advancements across NLP, information retrieval, computer vision, formal reasoning, and recommender systems, facilitating applications that require disambiguation, facet-aware retrieval, analogical matching, or polysemy resolution.

1. Conceptual Foundations and Problem Scope

Traditional similarity measures, such as cosine similarity over global word embeddings, fail to model context variability, asymmetry, syntax, or multi-faceted relations. The need for context-sensitive metrics originates in phenomena such as:

Word sense polysemy and context-driven semantic shift (Pilehvar et al., 2018, Armendariz et al., 2019).
Document similarity “with respect to” one or more aspects, e.g. background vs. methodology (Ostendorff, 2020).
Short-text paraphrase with minimal lexical overlap (0806.3787).
Visual similarity conditioned on a reference image (Liao et al., 2024).
Formal argument similarity dependent on support/claim separation and logical structure (David et al., 14 Apr 2026).

A context-sensitive similarity function is typically formalized as

$\mathrm{sim}_C(x, y; c)$

where $x, y$ are the objects compared and $c$ is the relevant context (linguistic, ontological, referential, or multimodal).

2. Methodological Taxonomy

Methods for context-sensitive similarity computation span the following paradigms:

A. Contextualized Embeddings.

Token/sentence/document representations adapt their vectorial form via context-dependent encoders (e.g., ELMo, BERT, RoBERTa) (Pilehvar et al., 2018, Al-Bataineh et al., 2019, Martinez-Gil, 2023, Sun et al., 2021). Pairwise similarity is computed via functions such as cosine or learned metrics.
For word-in-context tasks, models extract contextual token/sentence embeddings which are then compared (WiC, CoSimLex) (Pilehvar et al., 2018, Armendariz et al., 2019).
In unsupervised WSD, sense assignment is based on the minimal change to context embeddings when substituting synonym candidates (Martinez-Gil, 2023).

B. Feature-Driven or Hybrid Models.

Augmentation of base embedding similarity with external features—e.g., time, location for tweets, or bibliometric signals for documents—through additive or multiplicative integration schemes (Ostendorff, 2020, Hansel et al., 2018).
Contextual weights on features (e.g., ontological labels, document sections, graph nodes) modulate the importance of each component in the final similarity (Raeesi et al., 2014, Ostendorff, 2020, David et al., 14 Apr 2026).

C. Metric Learning and Contextualization.

Learning a context-dependent metric tensor (e.g., a positive semidefinite matrix $M$ for extended cosine similarity) regularizes or warps the embedding space to increase alignment with human judgments in a specific semantic domain or under a specific hypernym (Vos et al., 2022).
Cross-context transfer is validated via performance on multiple, independently labeled similarity benchmarks.

D. Ontology- and Structure-Based Methods.

Similarity in ontological trees incorporates both structural (path) and semantic (edge weights) proximity, with semantic weights estimated via normalized similarity scores derived from data (e.g., web co-occurrence via Google counts) (Raeesi et al., 2014).
Formal logic argument similarity relies on a compositional, multi-level aggregation—predicates, literals, clauses, and formulae—with contextual weighting at each level and satisfaction of axiomatic properties (David et al., 14 Apr 2026).

E. Multimodal and Triplet Approaches.

Contextualized ranking in image space: Context-sensitive similarity is operationalized as relative similarity to a reference, with ensemble models over localized neighborhoods to increase robustness against data sparsity (Liao et al., 2024).

3. Representative Formalisms and Algorithms

Contextualized Cosine and Metric Learning:

Let $\cos_M(u, v)$ denote the generalized cosine similarity with a learned metric $M = B^T B$ : $\cos_M(u, v) = \frac{u^T M v}{\sqrt{u^T M u}\sqrt{v^T M v}}$ where $B$ is optimized on context-specific datasets for maximal correlation with human scores (Vos et al., 2022).

Weighted Ontology-Tree Similarity:

For two context nodes $C_a, C_b$ , let the unique path be $P = (n_0, ..., n_k)$ . Each edge $x, y$ 0 receives weight $x, y$ 1 as

$x, y$ 2

where NGD is normalized Google distance. Similarity is then

$x, y$ 3

which decays with path length unless edges are semantically strong (Raeesi et al., 2014).

Context-Aware Semantic Similarity in WSD:

For target word $x, y$ 4 in context $x, y$ 5, with candidate synonyms $x, y$ 6, let $x, y$ 7 be the contextual encoder.

$x, y$ 8

Choose $x, y$ 9 as the sense minimizing meaning change (Martinez-Gil, 2023).

Multi-Level Argument Similarity (FOL):

For arguments $c$ 0 (as CNF sets of clauses), define at each level $c$ 1 a similarity function $c$ 2; recursively aggregate clause- and formula-level scores with context weights and best-match assignments. The argument-level similarity is

$c$ 3

with satisfaction of identity, symmetry, monotony, reinforcement, and other formal axioms (David et al., 14 Apr 2026).

4. Dataset Construction and Evaluation Protocols

Context-sensitive similarity research leverages a spectrum of dataset designs:

Word-in-Context Datasets:
- WiC: binary classification—does the target word have the same sense in each of two contexts? Gold labels achieved 80% annotator agreement; best BERT-based threshold models attain 65.5% (Pilehvar et al., 2018).
- CoSimLex: graded similarity scores ( $c$ 4) for word pairs in context; sensitive to both sense and salience variation; contextualized models range up to Spearman $c$ 5 (Armendariz et al., 2019).
Sentence and Document Similarity:
- Semantic Textual Similarity (STS): sentence pairs scored for degree of equivalence; context-based surrogate models achieve $c$ 6 on STSb (Sun et al., 2021).
- Knowledge-domain: contextual document similarity datasets designed for background/methodology/findings axes, with evaluation on nDCG, MAP, and click-through metrics (Ostendorff, 2020).
Image Similarity Benchmarks:
- Pooled triple-labeled 2AFC tasks (A:R vs. B:R); ensemble models outperform ResNet/CLIP/ViT stand-alone (Up to 84.7% accuracy) (Liao et al., 2024).

Standard evaluation metrics include Spearman rank correlation, Pearson $c$ 7, accuracy (for classification), mean absolute error, F1 for context classification, and ranking-based retrieval metrics (nDCG, MAP, MRR).

5. Applications Across Linguistic, Visual, and Logic Domains

Natural Language Processing:

Fine-grained paraphrase, duplicate detection, word sense disambiguation, and lexical substitute finding exploit context-sensitive similarity for robustness to polysemy, paraphrase, and style (Al-Bataineh et al., 2019, Martinez-Gil, 2023, 0806.3787).
Cultural analytics: measuring semantic drift or contestation of terms across genres or time (e.g., meaning of “revolution” in 18th-century French) (Zhou et al., 8 Feb 2025).

Information Retrieval and Recommender Systems:

Contextual document similarity enables faceted retrieval (e.g., similar findings but distinct methodologies) and diversified recommendation (Ostendorff, 2020).

Computer Vision:

Image retrieval systems adopt context-conditioned similarity—reframing the task as referential comparison—with mixture-of-experts ensembling for improved generalization (Liao et al., 2024).

Formal Argumentation and Logic:

Argument clustering, aggregation, and enthymeme decoding in FOL depend on multi-level, contextually weighted similarity metrics that integrate both symbolic and learned (SBERT-based) factors (David et al., 14 Apr 2026).

6. Strengths, Limitations, and Future Research Directions

Strengths:

Capture polysemy, sense salience, and aspectual similarity overlooked by global models.
Generalize across domains (text, vision, logic) via shared paradigm of context-conditioned scoring.
Enable analogical, facet-specific, or retrieval-by-example functionality.

Limitations:

Contextual datasets are expensive to create, especially for fine-grained, facet-localized, or multimodal settings.
Model architectures scaling to long documents or large image corpora present computational bottlenecks.
Predefined context taxonomies can impose rigid boundaries; learning contexts dynamically remains a challenge.

Active Research Frontiers:

End-to-end context selection/induction, eliminating hand-crafted context sets (Sun et al., 2021).
Interpretability and visualization of learned context metrics, especially metric tensors (Vos et al., 2022).
Efficient context integration for long-form and multimodal data (Ostendorff, 2020, Hansel et al., 2018).
Grounded and cross-lingual contextual similarity evaluation and adaptation (Armendariz et al., 2019, Pilehvar et al., 2018).
Axiomatic foundation for symbolic reasoning and its integration with neural/similarity-based approaches (David et al., 14 Apr 2026).

Context-sensitive similarity computation, underpinned by advances in contextual encoders, dynamic metric learning, and hybrid feature integration, is now an essential technology for computational linguistics, information retrieval, vision, and reasoning. Empirical work demonstrates its impact across lexical, sentence, document, and image understanding, while foundational studies rigorously formalize desirable properties and compositional structures.