How sense and meaning are shaped by experience has long been central to semantic research, particularly within linguistic, cognitive, and philosophical inquiry. Traditionally, semantic theory has debated whether meaning is best understood as a reflection of prototypical concepts, abstracted through repeated exposure, or as the product of compositional features learned through experience and encoded in language (Rosch, 1975; Carey, 2009). This debate touches on a deeper issue: Are linguistic meanings preserved in symbolic representations, or are they also grounded in somatic and experiential interactions with the world (Lakoff, 1987; Barsalou, 1999, 2008)?

Recent research in cognitive linguistics and embodied cognition increasingly emphasizes the importance of perceptual grounding in semantic understanding. The sensory experience - visual, tactile, auditory - plays a fundamental role in how an individual form conceptual categories and comprehend language (Glenberg and Kaschak, 2002; Gärdenfors, 2004). The smell of coffee or the glow of a sunrise are not merely experiential events; they help structure how we conceptualize related objects, events, and their linguistic referents. This experiential substrate informs the syntactic and semantic properties attributed to expressions across languages.

Languages often encode semantic properties through multiword expressions (MWEs), where meaning emerges not from individual words but from their inferential interplay between them. For example, the expression spill the beans demonstrates an idiomatic shift in which literal meanings yield to culturally learned experiential constructs: it signifies reveal a secret through embodied knowledge of spilling (accidental release) and beans (scattered objects), consistent with (Lakoff, 1987; Barsalou, 1999). Such expressions rely heavily on shared embodied knowledge and often resist modeling by traditional distributional or logic-based semantic frameworks (Nunberg et al., 1994; Jackendoff, 2002).

Given these challenges, our work is motivated by the need for a semantic annotation framework that can capture the dynamic and experiential dimensions of meaning across typologically diverse languages, where variation in syntactic structure often conceals deeper conceptual regularities. We propose a multidimensional approach grounded in a geometric model of meaning, where “geometric” denotes a structured, interpretable semantic space rather than opaque embeddings built from perceptual and cognitive dimensions. This model focuses on three primary semantic dimensions experiential, spatial, and temporal which we argue are foundational to meaning construction in typologically diverse languages.

To clarify these dimensions and their differences (noting potential overlaps, such as experiential aspects influencing spatial metaphors), we define them as follows:

Spatial dimension captures configuration, location, and distribution of entities or concepts, emphasizing physical or metaphorical positioning (e.g., shape, proximity, or orientation like “up” vs. “down”). This dimension draws from conceptual spaces where spatial topology structures meaning (Gärdenfors, 2004). Temporal dimension encompasses duration, sequence, and patterns over time, including progression, cyclicity, or frequency (e.g., “before/after” relations or ongoing events). It models dynamic aspects distinct from static space but often intersects with it in metaphors (Pala, 2023b; Arstila, 2016). Similarly, the experiential dimension reflects attributes grounded in bodily and sensory interactions, subdivided into functional (utility-based, e.g., an object's purpose), structural (organizational, e.g., composition), material (sensory features, e.g., texture or color), and qualitative (abstract/symbolic, e.g., emotional value or cultural associations). This is rooted in embodied cognition, where meaning emerges from perceptual simulations (Barsalou, 1999).

These dimensions interact to form multidimensional representations, accommodating semantic gradience. For example, in the metaphorical expression “rising tension,” the spatial dimension contributes upward movement (vertical configuration implying increase), the experiential dimension adds qualitative intensity (emotional or sensory strain from bodily tightness), and the temporal dimension conveys progression over time (escalation from low to high). This interplay, aligned with conceptual metaphor theory, describes how abstract experiences are structured through perceptual grounding (Lakoff, 1987). Each dimension reflects a network of conceptual relations that arise through perceptual engagement and are systematically encoded in language use (Evans, 2009).

To operationalize this framework, we introduce a set of POSI tags: 135 interpretable four-letter codes designed to annotate fine-grained semantic attributes within these dimensions. Unlike part-of-speech tagging or fixed ontological categories, POSI tags capture dynamic, context-activated semantic properties rooted in cognitive experience. This makes the framework particularly effective for annotating MWEs, whose meanings often defy conventional syntactic or lexical classification.

We evaluated our approach by annotating MWEs in four structurally and typologically diverse languages. This multilingual evaluation supports our hypothesis that perceptually grounded dimensions of meaning are both cross-linguistically robust and computationally tractable. The proposed model offers a nuanced geometric representation of meaning that accommodates semantic gradience, overlap, and context sensitivity.

By integrating formal semantic theory with embodied cognition and multilingual linguistic data, this work provides a novel toolset for semantic analysis and offers theoretical insights into how experience shapes linguistic representation. The following sections describe the design of the geometric model, the structure, and rationale behind the POSI tags, and our methodology for multilingual annotation.

The geometric model of experiences serves as a metaphorical framework to analyze the complexity of perception, interpretation, and navigation of subjective realities. By analogizing geometric principles to experiential dimensions, this model extends spatial and dimensional concepts to explore how individuals construct meaning across emotional, sensory, and cognitive domains. Semantic properties of objects or representations are structured into three dimensions - spatial, temporal, and experiential proposed by Pala (2023b). Experiential properties further decompose into functional, material, structural, and qualitative aspects, which manifest as perceptual modality-oriented expressions in linguistic discourse (Pala and Shalu, 2025). This formalization enables systematic analysis of how inferential processes synthesis sensory, cognitive, and symbolic information into coherent representations.

Experiential properties hold particular significance due to their grounding in bodily sensations, which mediate both real-world interactions and imagined scenarios (Pala, 2023a). For example, experiencing a red apple involves integrating visual (redness), tactile (material texture), gustatory (sweetness or sourness), olfactory (smell), and cognitive (shape, temperature) inputs. Such multimodal correlations underpin the formation of representational content, as the synthesis of these properties shapes interpretable meaning. In figurative language, subjective experience drives inferential processes. The expression “John has a heart of gold,” for example, invokes qualitative attributes of gold (warmth, value) alongside spatiotemporal cues (“has a heart of”). This maps to the paraphrase “John is kind and thoughtful,” demonstrating how experiential properties guide metaphorical reasoning. To transition from conceptual framing to formal implementation, we introduce a structured semantic space that mirrors how experiential, spatial, and temporal features are encoded linguistically. This space underlies our POSI tagging system, where each four-letter code maps to salient coordinates along these dimensions. The term “geometric” is retained over “vector representations” to emphasize the structured, interpretable nature of this representational space. While grounded in ℝⁿ vectors, the model transcends numerical embedding by encoding dimensions with semantic salience. Spatial, temporal, and experiential properties are mapped to subspaces that preserve perceptual and cognitive relationships. This allows modeling not only proximity but also metaphorical projection and analogical reasoning-processes aligned with embodied cognition and conceptual metaphor theory. For example, expressions like “rising tension” reflect spatialized conceptualizations of experience. Structuring semantic spaces geometrically ensures alignment with reasoning patterns (e.g., blending, interpolation), bridging formal semantics and cognitive processes. The framework provides a foundation for semantic annotation and computational applications. By formally encoding experiential properties in geometric vectors, it enables both theoretical analysis and practical tools in computational linguistics and cognitive science, while maintaining cognitive plausibility through its emphasis on spatial organization and cross-modal integration.

As an initial step, we manually analysed 100 English verb-noun idiomatic expressions selected from the VNC-Tokens dataset (Cook et al., 2008) and annotated with the proposed POSI tags. We focused on idiomatic expressions because we think linguistic expressions with figurative sense will be a suitable test case for validating the effectiveness of the geometric model and illustrating its nuances. Note that the geometric model used in this work posits that geometric properties of a figurative expression can be derived from the geometric properties of its constituent parts.

While we claim compositionality, we recognize that many figurative units are non-compositional: German “jmdm. einen Bären aufbinden” (“tell tall tales”), English “kick the bucket” (“die”) and Bangla “Ghorar dim” (“horse's egg” = “something impossible”) cannot be built from the spatial, material or functional properties of their parts alone. For such cases, the POSI fulfiller tags host the cultural/historical knowledge that bridges literal components and idiomatic sense, instead of forcing a geometric sum that never existed (Pala et al., 2025; DWDS, 2024; Nunberg et al., 1994).

We are not limiting the scope of our proposed model and annotation framework to idiomatic expressions and expressions from the English language. To validate the cross-structural applicability, we have annotated sets of transitive sentences and intransitive sentences from the British National Corpus (BNC) (Burnard, 2007; Consortium, 2007). To validate cross-lingual applicability, we selected three non-English languages, Punjabi and Bangla from the Indo-Aryan language family and Malayalam from the Dravidian language family. Then we have annotated five sample idiomatic expressions from each of these three non-English languages. We used GyanNidhi Multilingual Parallel Corpus and other linguistic resources in Indian languages (Consortium, 2008; Choudhary, 2019; Chaudhury et al., 2008; Arora et al., 2003; Gupta and Pala, 2012; Agrawal, 2008) for selecting expressions for the annotation. Annotated sample of an idiomatic expression, “find foot” is included in Appendix 1 and we will publicly release our entire annotated data.

In this paper, we introduced a geometric semantic model that encodes meaning as interpretable vectors along spatial, temporal and experiential dimensions, together with POSI - 135 four-letter tags that annotate fine-grained, perceptually grounded semantic properties. A multilingual evaluation on English, Punjabi, Bangla, and Malayalam idioms, transitive/intransitive clauses and other MWEs showed that the scheme is cross-linguistically robust and cognitively plausible. By formalizing experiential, spatial, and temporal features that traditional models treat as peripheral, the framework bridges embodied-cognition insights and computational semantics, offering a unified account of how bodily experience, cultural knowledge, and linguistic structure jointly construct meaning, especially in non-compositional or metaphorical expressions.

Each POSI tag maps to a transparent region of the geometric space, so the resource can be plugged directly into downstream NLP pipelines: the tags supply explainable, embodied features to semantic parsers, metaphor detectors and machine-translation systems, while the fulfiller dimension supplies cultural or historical knowledge whenever compositionality fails.

In the future work, we will (i) automate POSI tagging via supervised learning, (ii) embed the tags into existing semantic knowledge bases, and (iii) extend the inventory to additional languages and genres. And we will undertake a quantitative evaluation-comparing POSI-augmented models against strong baselines on tasks such as semantic parsing to demonstrate the concrete effectiveness.