A Textual Semantic Analysis Framework Integrating Geographic Metaphors and GIS-Based Spatial Analysis Methods

Abstract

Geographic information systems (GISs) have shown considerable promise in enhancing textual semantic analysis. Current textual semantic analysis methods face significant limitations in accurately delineating semantic boundaries, identifying semantic clustering patterns, and representing knowledge evolution. To address these issues, this study proposes a framework that innovatively introduces GIS methods into textual semantic analysis and aligns them with the conceptual foundation of geographical metaphor theory. Specifically, word embedding models are employed to endow semantic primitives with comprehensive, high-dimensional semantic representations. GIS methods and geographical metaphors are subsequently utilized to project both semantic primitives and their relationships into a low-dimensional geospatial analog, thereby constructing a semantic space model that facilitates accurate delineation of semantic boundaries. On the basis of this model, spatial correlation measurements are adopted to reveal underlying semantic patterns, while knowledge evolution is represented using ArcGIS 10.7-based visualization techniques. Experiments on social media data validate the effectiveness of the framework in semantic boundary delineation and clustering pattern identification. Moreover, the framework supports dynamic three-dimensional visualization of topic evolution. Importantly, by employing specialized visualization methods, the proposed framework enables the intuitive representation of semantic symmetry and asymmetry within semantic spaces.

1. Introduction

Geographic information systems (GISs) have proven particularly effective in spatial data analysis across diverse disciplines, such as urban planning, environmental monitoring, historical geography, and public health. Characterized by their unique capabilities in spatial positioning, data integration, analysis, and visualization, GIS methods have evolved into critical tools for interdisciplinary research [1,2]. In the era of rapidly expanding digital information, GIS methods provide innovative means to analyze complex spatial data, empowering researchers to decode spatial relationships with enhanced accuracy [1,2,3,4]. In particular, in the field of textual semantic analysis, GIS methods were introduced in the early 2000s, yielding a range of notable achievements [1,5,6]. Numerous studies have utilized GIS methods to perform overlay analyses of textual and geospatial data, facilitating not only the extraction of deep knowledge but also the identification of semantic patterns and knowledge evolution [5,7].

Traditional methods in textual semantic analysis are largely based on co-occurrence statistics and feature extraction, which may not adequately clarify the semantic boundaries of individual semantic primitives (referring to fundamental units of meaning that serve as the basic analytical elements within a semantic space, such as words, phrases, or concepts) [8,9,10]. This limitation restricts their ability to identify both clustering patterns—such as symmetry or asymmetry—among semantic units and knowledge evolution processes, thereby undermining the effectiveness of semantic analysis and the accuracy of trend prediction in knowledge evolution [11,12,13]. Previous studies have noted that higher accuracy of semantic boundary delineation can be achieved by mapping semantic primitives as point features into a semantic space where semantic relationships are quantified by spatial distances [14,15,16]. However, such approaches face significant challenges due to the high dimensionality and abstract nature of the semantic space [17,18,19]. In recent years, drawing on the theory of geographic metaphors, researchers have attempted to metaphorically model high-dimensional semantic space as a low-dimensional geospatial analog [20,21]. These attempts successfully reduce the cognitive burden associated with interpreting semantic space, demonstrating the potential of GIS methods for more accurate delineation of semantic boundaries. Notably, the powerful capabilities of GIS methods in spatial analysis and visualization can also be utilized for identifying semantic clustering patterns and representing knowledge evolution.

Given the above, this study proposes a novel framework that integrates both geographic metaphors and GIS methods involving spatial analysis and visualization into textual semantic analysis. The framework endows semantic primitives with comprehensive, high-dimensional semantic representations using word embedding models. Guided by the theory of geographic metaphors and through a tailored dimensionality reduction technique, these representations are subsequently transformed into a geospatial-analogous semantic model. On the basis of this model, analyses of spatial correlation are combined with two-dimensional (2D) visualization to identify semantic clustering patterns. In parallel, three-dimensional (3D) visualization techniques are applied to represent knowledge evolution in the form of a semantic landscape. This landscape serves as a metaphorical, topographic representation where the distribution and intensity of semantic primitives are mapped as elevations, revealing structural and thematic dynamics within the semantic space [22,23,24].

Within the proposed framework, GIS methods act as the methodological backbone, underpinning the entire process—from semantic boundary delineation and clustering pattern identification to the interpretation of knowledge evolution. Hence, this study addresses three key questions: (1) Can GIS methods be effectively applied to accurately delineate semantic boundaries, and if so, in what way? (2) Can the GIS method be utilized to identify semantic patterns, and if so, how? (3) Can the GIS method be employed to interpret knowledge evolution? If so, how can it be achieved?

The remainder of this paper is organized as follows: The next section reviews related work, followed by a description of the proposed framework. Experiments conducted on real-world data are subsequently presented, and their results are comprehensively analyzed. The findings are then discussed in depth, and finally, the paper concludes by summarizing the results and suggesting directions for future research and potential trends.

2. Related Work

2.1. Geographic Space and Semantic Space

In geography, geographic space is defined as a collection of entities associated with spatial reference information, comprising objectively existing and interrelated geographic elements. This space is characterized by five principal attributes: diversity, distribution, correlation, variability, and accessibility [25,26,27]. Diversity reflects heterogeneity in the quantity, type, and organization of geographic elements. Distribution refers to the spatial arrangement of natural and human features, whose patterns may exhibit features such as symmetry or asymmetry, reflecting different degrees of spatial balance. Correlation refers to dependencies among spatial entities, which often exhibit similar or covarying attributes. Variability emphasizes that spatial phenomena differ across locations and scales. Accessibility denotes that each geographic entity has both absolute (e.g., latitude, longitude, elevation) and relative positions (e.g., distance or orientation to other objects), forming the basis for spatial relationships. Geographic entities exhibit two essential attributes: positional and relational. Positional attributes specify location, typically represented by spatial primitives such as points, lines, and polygons. Relational attributes describe how entities connect on the basis of geometric properties (e.g., distance, direction, proximity) or statistical associations (e.g., spatial correlation, dependence) [25,28]. The most fundamental spatial relationship is the binary relationship, in which two entities are considered related if they are geographically close; otherwise, they are unrelated [27,29].

In contrast, semantic space is an abstract conceptual domain composed of semantic primitives and their interrelationships, offering a mathematical framework for representing linguistic meaning and its underlying connections [30]. The concept emerged from two core challenges in natural language processing: word mismatch (same meaning, different expressions) and semantic ambiguity (same word, multiple meanings). Its earliest formalization dates back to Collins and Quillian (1969), who proposed a hierarchical semantic network to structure relationships between primitives and their attributes [31]. Later, Collins and Loftus (1975) introduced the notion of semantic distance, which quantifies conceptual similarity via network proximity [32]. Over subsequent decades, various classical methods have emerged, such as latent semantic analysis (LSA) [33], hyperspace analog to language (HAL) [34], and pointwise mutual information (PMI) [35]. While these classical models offer advanced semantic representations, they often overlook relational structures beyond co-occurrence statistics, such as those involving semantic symmetry, thereby limiting their ability to capture semantic topographies or spatiotemporal dynamics. With rapid advances in artificial intelligence, semantic space research has transformed. Modern algorithms reconstruct contextual word relationships and represent primitives as points in high-dimensional space, typically through unsupervised learning across hundreds of dimensions. This allows the precise quantification of semantic differences on the basis of distance while capturing the connotations, structures, and distributional properties of semantic primitives [11,36,37]. Modern embedding models improve semantic resolution but still lack an intuitive means to visualize and interpret semantic relationships at macro scales, particularly when modeling evolution and clustering over time.

Despite originating from distinct disciplines, geographic space and semantic space share several structural similarities. In geographic space, most attribute information can be extracted through spatial overlay analysis [25,28]. Similarly, in semantic space, each unit (word or phrase) is positioned in a high-dimensional vector space, where its relative position encodes its meaning and relationships with other units [12,37,38]. These similarities make it possible to establish mapping relationships between these two spaces, especially in terms of structure, clusters, hierarchy, and spatial correlation. Establishing such connections provides a novel way to address complex problems in semantic analysis by leveraging spatial modeling techniques originally developed for geographic data.

On the other hand, in recent years, geographic metaphors have been increasingly employed in semantic analysis. Rooted in cognitive linguistics, this approach posits that abstract concepts can be understood via concrete spatial experiences, as exemplified by the concept of a “semantic landscape”. In this metaphor, a semantic landscape, composed of peaks (high-frequency concepts), valleys (low-frequency or peripheral concepts), clusters, and boundaries, provides an intuitive and visual framework for modeling meaning. By employing geographic metaphors, researchers can apply mature GIS methods, such as spatial clustering, spatial correlation, and topological mapping, to the exploration of semantic relationships that traditional methods fail to reveal. In addition, geographic metaphors become especially useful when semantic systems exhibit complexity, multiscale interactions, and temporal dynamics—characteristics commonly seen in both linguistic and spatial data. Representing semantic units as spatial analogs enables dynamic modeling, intuitive visualization, and relational analysis.

2.2. Spatial Correlation and Semantic Correlation

The first law of geography states, “everything is related to everything else, but near things are more related to each other”. This principle highlights the intrinsic spatial dependency among entities, elements, or phenomena within geographic space. Rather than existing in isolation, geographic entities interact with and influence one another, with spatial proximity often leading to stronger correlations in attributes or characteristics. For example, temperatures in geographically adjacent areas tend to exhibit high correlations. Figure 1 shows the temperature distributions across the United States on a specific day in June 2021, with the temperature exhibiting a distinct asymmetrical distribution across the entire country. Taking Arizona as an example, location A is closer to location B than to location C (Figure 1), meaning that the temperature at location B is more strongly correlated with that at location A than with that at location C. Moreover, the inherent intuitiveness of geographic space facilitates the identification of spatial entities, their attributes, and their evolutionary patterns. This spatial awareness enhances our ability to analyze and interpret geographic relationships and processes. Currently, spatial correlation analysis plays a vital role in a wide range of fields, including environmental science, epidemiology, urban planning, and transportation system analysis.

Semantic correlation is a fundamental concept in natural language processing (NLP) and information retrieval and describes the degree of association between two or more semantic primitives (e.g., words, phrases, or documents) at the semantic level. Unlike simple word matching, semantic correlation captures deeper meanings and contextual relationships within textual data [11,39,40]. Mainstream methods such as word–vector models primarily learn word semantics through neural networks, mapping words into high-dimensional vector spaces where semantic correlation is assessed via vector similarity metrics [11,41,42]. In recent years, deep learning-based models (e.g., bidirectional encoder representations from transformers (BERT)) have significantly enhanced the accuracy of semantic correlation calculations by leveraging large-scale pretraining [10,43]. As a result, semantic correlation analysis has been widely applied across NLP tasks such as machine translation, text mining, and question-answering systems [11,41,43]. More importantly, recent studies indicate that semantic correlations within domain-specific texts can reveal key insights into the evolution of thematic trends in various fields [44,45]. For example, in the domain of information retrieval, information asymmetry often leads users to utilize search engines to access the information they need. Figure 2 illustrates this dynamic retrieval process: a user initially accesses paper P1 but later transitions to reading P4, which is more relevant and up-to-date. P4 is subsequently cited by web article W4, leading them to explore W4, thereby accelerating the dissemination of both P4 and W4. This retrieval path highlights the role of semantic correlation in knowledge dissemination and thematic evolution.

In summary, spatial correlation and semantic correlation share several key characteristics. First, they both rely on quantitative measures, such as Moran’s I or spatial distance for spatial correlation, and cosine similarity or word embedding models for semantic correlation. Second, they are both subject to external influences. Spatial correlation may be affected by spatial scale, distribution patterns, and other geographic characteristics, whereas semantic correlation can be influenced by part of speech, context, and text type [46]. Moreover, both spatial and semantic correlation methodologies are versatile and domain-independent, supporting the analysis of continuous, discrete, or mixed data in various physical and social science fields [11,47]. Spatial correlation analysis has been widely used not only for analyzing continuous physical phenomena (e.g., temperature) but also for discrete or categorical social processes, such as crime clustering, population migration, and the spread of infectious diseases [46]. Likewise, semantic correlation analysis extends beyond discrete social data like citations, finding application in scientific literature, legal documents, and even biological data annotations [20,48,49].

2.3. Current Challenges and the Contributions of This Work

Though recent studies have made some progress by applying geographic metaphors and GIS methods to textual semantic analysis, several key challenges remain to be addressed.

(1)

Semantic boundary delineation. Although semantic distances between primitives can be computed in high-dimensional semantic space, there remains a lack of systematic standards for defining the semantic scope of individual primitives. Consequently, existing studies primarily rely on subjective thresholds or manual heuristics to delineate semantic boundaries, leading to instability and limited generalizability. Moreover, mainstream semantic modeling approaches are developed based on models lacking explicit mechanisms for boundary delineation, such as bag-of-words models or topic models [6,13,50], which makes it difficult for them to distinguish core semantic clusters from transitional zones.

(2)

Semantic clustering pattern identification. Previous studies predominantly focus on measuring semantic correlation or representing co-occurrence networks, leaving the discovery of latent semantic patterns largely underexplored. Furthermore, current methods mainly adopt static clustering strategies, failing to capture the symmetric, hierarchical, and context-dependent nature of semantic associations [10,12,51]. The absence of well-defined semantic boundaries also hampers the identification of underlying semantic clustering patterns.

(3)

Knowledge evolution representation. Current methods for visualizing knowledge evolution mainly rely on time-series charts or 2D planar layouts, which are insufficient for clearly illustrating dynamic processes such as semantic expansion and migration. This limitation becomes more pronounced when handling cross-temporal, multimodal, and heterogeneous data, where 2D visualizations struggle to represent the complex multidimensional dynamics of semantic evolution. As a result, critical details regarding semantic relationship changes and structural reconfigurations are often omitted [52,53,54,55]. These limitations reduce the interpretability of visualizations and impair the analytical utility of such representations in understanding knowledge evolution.

To address these challenges, this study proposes an integrated textual semantic analysis framework that combines geographic metaphor theory with GIS methods. The framework metaphorically transforms high-dimensional semantic space into a geospatial analog, allowing for precise delineation of semantic boundaries, as well as multiscale identification of clustering patterns using multiple spatial correlation techniques. Furthermore, by leveraging ArcGIS, the framework constructs a 3D semantic landscape model that dynamically visualizes semantic clustering and diffusion processes, thereby enhancing the interpretability of knowledge evolution.

3. Method

An overview of the proposed framework is illustrated in Figure 3. This approach achieves accurate delineation of semantic boundaries by modeling the semantic space as a geospatial analog. Semantic correlation analyses are subsequently conducted within the geospatial analog to identify semantic patterns. Finally, knowledge evolution is represented using ArcGIS-based visualization techniques in a form analogous to urban morphological dynamics. The details of how the framework models the semantic space, performs semantic correlation analyses, and represents knowledge evolution are presented below.

3.1. Semantic Space Modeling

3.1.1. Workflow

First, text data are processed using the Jieba 0.39 segmentation tool for tokenization and preprocessing.

Second, a Word2Vec-based word embedding model is constructed using the skip-gram algorithm, with a window size of 10 and a vector dimension of 400, to capture the semantic representations of inferiority-related expressions.

Third, the semantic frequency–semantic active index (SF–SAI) algorithm is applied to extract semantic primitives and compute both their absolute frequencies and the frequencies of semantically related terms.

Fourth, given that the extracted semantic primitives are high-dimensional vectors, a nonlinear dimensionality reduction technique—t-distributed stochastic neighbor embedding (t-SNE)—is employed to project them into a low-dimensional space for visualization (

Table 1. Sample of word vectors reduced to two dimensions.

ID	Semantic Primitive	X	Y	English
1	傲气	−0.859604166	1.984212902	Arrogant
2	倔强	0.643495166	−1.313592229	Obstinate
3	天生	−0.573996029	−1.886607475	Innate
4	偏执	1.278016735	−0.601750412	Stubborn
5	强势	0.515507556	1.197673322	Mighty
6	功利	−1.143747358	0.450880872	Benthamism
7	骗子	−1.37262951	0.895490501	Fraud
8	无理	−0.612900011	−1.342502749	Unreasonable
9	骄傲	−0.122557241	0.501880319	Pride
10	自私	0.854270125	−0.839828016	Selfishness

).

Fifth, guided by the theory of geographic metaphors, a semantic space analogous to geographic space is constructed. Specifically, the data (shown in Table 1) are imported into ArcGIS, where each semantic primitive is represented as a point on the 2D visualization plane (Figure 4). These points are then treated as “seed points” for their respective regions, and geometric divisions are performed using the nearest neighbor principle. Through this partitioning process, a geospatial-analogous model for semantic analysis is established.

3.1.2. Dimensionality Reduction

As mentioned earlier, the framework employs t-SNE to reduce the dimensionality of semantic vectors. As a classical manifold learning algorithm, t-SNE optimizes the divergence between probability distributions in high-dimensional and low-dimensional spaces using the Kullback–Leibler (KL) divergence, thereby preserving the topological structure of the original semantic space during projection [46,47]. The resulting low-dimensional representation is further formalized through a semantic proximity matrix, which captures latent associations between primitives and offers a mathematically interpretable basis for semantic correlation analysis [49].

The use of t-SNE is motivated by three key considerations. First, textual semantic data are typically characterized by high dimensionality and nonlinear structure. These characteristics are difficult to capture effectively using traditional linear dimensionality reduction techniques such as principal component analysis (PCA) and multidimensional scaling (MDS). Unlike those techniques, t-SNE is specifically designed to handle high-dimensional, nonlinear data. Second, t-SNE excels at maintaining local similarities and neighborhood structures, enabling semantic units to form compact clusters in the reduced 2D space while clearly defining the boundaries and transitional zones between semantic groupings. This minimizes structural distortion or ambiguity during the dimensionality reduction process. Third, the output of t-SNE is intuitively interpretable, which facilitates follow-up spatialized semantic analysis.

3.2. Semantic Correlation Analysis

As presented in Figure 5, the framework introduces classical spatial correlation measures—Moran’s I, Geary’s C, and Getis-Ord’s G—into semantic correlation analysis. The introduction is based on the fact that once high-dimensional semantic vectors are projected into a 2D space, the positional relationships among semantic primitives exhibit mathematical structures analogous to adjacency and distance in geographic space [56,57,58]. In addition, once semantic correlation is metaphorically represented as spatial proximity, a geospatial-analogous semantic field is formed, which involves clusters, transitional zones, and expansion boundaries [59,60]. Moran’s I can be used to assess the overall tendency of semantic primitives to cluster, determining whether the observed patterns are statistically significant. In contrast, the Geary C index, which is more sensitive to local variation, facilitates the detection of semantic boundaries, transitional regions, and areas of abrupt semantic change. Moreover, the Getis-Ord G index enables the precise identification of semantic hotspots and their expansion boundaries, thereby capturing the spatial concentration of semantic intensity. These measures can quantitatively assess the spatial structure of semantic relationships, facilitating semantic pattern identification and knowledge evolution representation [61,62,63].

3.2.1. Global Semantic Correlations

The formula for calculating the $\mathrm{g}\mathrm{l}\mathrm{o}\mathrm{b}\mathrm{a}\mathrm{l} \mathrm{G}\mathrm{e}\mathrm{t}\mathrm{i}\mathrm{s}-\mathrm{O}\mathrm{r}\mathrm{d} \mathrm{G}$ index (G) is as follows:

$$G={\displaystyle \frac{{\sum }_{i=1}^n{\sum }_{j=1}^n{w}_{ij}{x}_i{x}_j}{{\sum }_{i=1}^n{x}_i^2}}$$

(1)

where $G\approx E\left[G\right]$ indicates that no clustering of high values (hot spots) or low values (cold spots) occurs in the spatial regions representing semantic primitives. However, if $G>E\left[G\right]$, significant clustering of high values (hot spots) occurs in the spatial regions representing semantic primitives. Conversely, when $G<E\left[G\right]$, significant clustering of low values (cold spots) occurs in these regions.

The global Moran’s I index (I) ranges from −1 to 1. When $I\in [-\mathrm{1,0})$, the spatial regions representing semantic primitives exhibit negative spatial correlation. When $I=0$, there is no spatial correlation. Conversely, when $I\in \left(\mathrm{0,1}\right]$, it signifies positive spatial correlation. The formula for calculating the $\mathrm{g}\mathrm{l}\mathrm{o}\mathrm{b}\mathrm{a}\mathrm{l} \mathrm{M}\mathrm{o}\mathrm{r}\mathrm{a}\mathrm{n}\mathrm{’}\mathrm{s} \mathrm{I}$ index is as follows:

$$I={\displaystyle \frac{n{\sum }_{i=1}^n{\sum }_{j=1}^n{w}_{ij}\left(x_i-\stackrel{-}{x}\right)\left(x_j-\stackrel{-}{x}\right)}{{\sum }_{i=1}^n{{\sum }_{j=1}^n{w}_{ij}{\sum }_{i=1}^n\left(x_i-\stackrel{-}{x}\right)}^2}}$$

(2)

The $\mathrm{g}\mathrm{l}\mathrm{o}\mathrm{b}\mathrm{a}\mathrm{l} \mathrm{G}\mathrm{e}\mathrm{a}\mathrm{r}\mathrm{y}\mathrm{’}\mathrm{s} \mathrm{C}$ index (C) ranges from 0 to 2. When $\mathrm{C}=1$, there is no spatial correlation, meaning that the spatial regions representing semantic primitives are randomly distributed. When $C\in \left[\mathrm{0,1}\right]$, it suggests positive spatial correlation, where neighboring regions have similar observed values, resulting in clustering or pattern similarity. Conversely, when $C\in \left[\mathrm{1,2}\right]$, it indicates negative spatial correlation, where neighboring regions exhibit substantial differences in observed values, resulting in strong spatial heterogeneity. The calculation formula for the $\mathrm{g}\mathrm{l}\mathrm{o}\mathrm{b}\mathrm{a}\mathrm{l} \mathrm{G}\mathrm{e}\mathrm{a}\mathrm{r}\mathrm{y}\mathrm{’}\mathrm{s} \mathrm{C}$ index (C) is as follows:

$$C={\displaystyle \frac{(n-1){\sum }_{i=1}^n{\sum }_{j=1}^n{w}_{ij}{(x_i-x_j)}^2}{2({\sum }_{i=1}^n{\sum }_{j=1}^n{w}_{ij}){\sum }_{i=1}^{i=n}{(x_i-x)}^2}}$$

(3)

Furthermore, in Formulas (1)–(3), $n$ represents the number of semantic primitives in the database;$x_i \mathrm{a}\mathrm{n}\mathrm{d} x_j$ denote the occurrence frequencies of semantic primitives $i \mathrm{a}\mathrm{n}\mathrm{d} j$, respectively; and $w_{ij}$ reflects their spatial relationship (adjacency).

3.2.2. Local Semantic Correlations

The formula for the local $\mathrm{G}\mathrm{e}\mathrm{t}\mathrm{i}\mathrm{s}-\mathrm{O}\mathrm{r}\mathrm{d} {\mathrm{G}}_{\mathrm{i}}^{\mathrm{*}}$ index $\left(G_i^{\mathrm{*}}\right)$ for semantic primitive $i$ is given by the following:

$$G_i^{\mathrm{*}}={\displaystyle \frac{{\sum }_{j=1}^n{w}_{ij}{x}_j}{{\sum }_{j=1}^n{x}_j}}$$

(4)

To assess the statistical significance of $G_i^{\mathrm{*}}$, the standardized statistic ${Z(G}_i^{\mathrm{*}})$ is calculated as follows:

$${\mathrm{Z}(\mathrm{G}}_{\mathrm{i}}^{*}) ={\displaystyle \frac{ {\sum }_{\mathrm{j}=1}^{\mathrm{n}}{\mathrm{w}}_{\mathrm{ij}}({\mathrm{x}}_{\mathrm{j}} - \stackrel{\mathrm{-}}{\mathrm{x}}) }{\mathrm{S}\sqrt{\frac{\mathrm{n} {\sum }_{\mathrm{j}=1}^{\mathrm{n}}{\mathrm{w}}_{\mathrm{ij}}^2- {\left({\sum }_{\mathrm{j}=1}^{\mathrm{n}}{\mathrm{w}}_{\mathrm{ij}}\right)}^2 }{\mathrm{n} - 1}}}}$$

(5)

where $G_i^{\mathrm{*}}$ is the $Getis-Ord G_i^{\mathrm{*}}$ statistic for semantic primitive $i$, and $S$ is the standard deviation. If ${Z(G}_i^{\mathrm{*}})>0$, it indicates a significant clustering of high values in the region where the semantic primitive $i$ is located, as well as in the regions of its adjacent semantic primitives. If ${Z(G}_i^{\mathrm{*}})<0$, it indicates significant clustering of low values in the region of semantic primitive $i$ and its neighboring regions. If ${Z(G}_i^{\mathrm{*}})\approx 0$, there is no significant spatial clustering, with the data approaching a random spatial distribution.

The local $\mathrm{M}\mathrm{o}\mathrm{r}\mathrm{a}\mathrm{n}\mathrm{’}\mathrm{s} \mathrm{I} \mathrm{i}\mathrm{n}\mathrm{d}\mathrm{e}\mathrm{x} {(I}_i)$ for semantic primitive $i$ is calculated as follows:

$$I_i={\displaystyle \frac{(x_i-\stackrel{-}{x})}{S^2}}{\sum }_{j=1}^n{w}_{ij}(x_j-\stackrel{-}{x})$$

(6)

where $S^2={\displaystyle \frac{1}{n}}{\sum }_{i=1}^n{(x_i-\overline{x})}^2$. On the basis of the value of $I_i$, local spatial correlation is categorized into four types: high–high (HH), low–low (LL), high–low (HL), and low–high (LH). Specifically, the types are further defined as follows: high–high indicates that $I_i\in \left(\mathrm{0,1}\right] \mathrm{a}\mathrm{n}\mathrm{d} Z\left(I_i\right)\in (1.96,+\mathrm{\infty });$ low–low signifies that $I_i\in \left(\mathrm{0,1}\right] \mathrm{a}\mathrm{n}\mathrm{d} Z\left(I_i\right)\in \left(-\mathrm{\infty },-1.96\right);$ high–low indicates that ${\mathrm{I}}_{\mathrm{i}}\in \left[-\mathrm{1,0}\right) \mathrm{a}\mathrm{n}\mathrm{d} \mathrm{Z}\left({\mathrm{I}}_{\mathrm{i}}\right)\in \left(-\mathrm{\infty }, -1.96\right);$ and low–high indicates that $I_i\in [-\mathrm{1,0}) \mathrm{a}\mathrm{n}\mathrm{d} Z\left(I_i\right)\in (1.96, +\mathrm{\infty })$.

The local $\mathrm{G}\mathrm{e}\mathrm{a}\mathrm{r}\mathrm{y}\mathrm{’}\mathrm{s} \mathrm{C} \mathrm{i}\mathrm{n}\mathrm{d}\mathrm{e}\mathrm{x} \left(C_i\right)$ for the semantic primitive $i$ is calculated as follows:

$$C_i={\displaystyle \frac{{\sum }_{j=1}^n{w}_{ij}{(x_i-x_j)}^2}{2{\sum }_{j=1}^n{(x_j-\overline{x})}^2}}$$

(7)

If $C_i>1$, a negative spatial correlation exists between the region representing semantic primitive $i$ and its neighboring regions. If $C_i<1$, a positive spatial correlation exists. When $C_i=1$, there is no local spatial correlation.

3.3. Knowledge Evolution Representation

The framework represents knowledge evolution using ArcGIS-based techniques.

Specifically, on the basis of the constructed geospatial-analogous semantic space, the frequency of each semantic primitive is mapped to the “elevation” of its corresponding semantic region, thereby generating a 3D semantic landscape with topographic characteristics. This 3D representation draws on conceptual metaphor theory from cognitive linguistics as well as geographic metaphor theory, wherein abstract semantic intensity is metaphorically rendered as terrain elevation [64]. This representation visualizes both the relative importance and the temporal evolution of semantic primitives within the semantic system.

Unlike traditional representation methods, which are normally 2D, the proposed method enables visual and quantifiable representation by generating a semantic landscape. It allows researchers to clearly identify “semantic highlands” (semantic primitives with high frequency, strong clustering, and central thematic relevance) as well as “semantic lowlands” (semantic primitives with low frequency, weak clustering, or thematic relevance), thereby uncovering evolutionary stages, shifting centers of cognitive focus, and dynamic transformation patterns of specific research themes over time. Moreover, considering the temporal evolution of semantic hotspots, this method further maps the expansion direction and rate of semantic primitives over time as “semantic expansion trajectories”, visually representing the developmental pathways and growth rhythms of the research theme. This representation strategy not only enhances the interpretability of the knowledge evolution process but also provides a novel perspective and analytical foundation for exploring knowledge diffusion paths and the mechanisms underlying topic emergence [65,66].

4. Experiment

4.1. Data

In this study, a total of 1.2 million Sina Weibo posts related to the theme of “inferiority”, published between 1 January 2011 and 31 December 2017, were collected. This dataset was selected for the following reasons.

Sina Weibo is one of the most influential social media platforms in China, with a massive and diverse user base spanning nearly all age groups and educational backgrounds. The platform hosts hundreds of millions of monthly active users and generates over a billion posts daily [67,68]. Its content spans a broad spectrum of topics, including domestic and international news, public opinion, social issues, and individual-level expressions of psychology and emotion [69,70,71]. Notably, the informal, unstructured, and non-standardized nature of user-generated content on Sina Weibo provides a rich textual corpus for semantic research and serves as a suitable testbed for evaluating the robustness of textual semantic analysis frameworks [72].

Second, inferiority, as a common psychological issue, is widely discussed across social media. Posts related to this theme exhibit semantically rich and diverse linguistic expressions, reflecting complex emotional and cognitive variations at both the individual and societal levels [73,74,75,76]. These characteristics make the dataset large-scale, heterogeneous, and context-sensitive, thereby supporting a comprehensive evaluation of the proposed framework and enabling in-depth exploration of evolving semantic relationships.

All texts used in this study were collected from publicly accessible Sina Weibo posts using keyword-based retrieval methods. No authentication-protected content, personal identifiers, or location metadata were accessed throughout the study. The dataset was fully anonymized and processed solely at the semantic level, with no attempt to infer or profile individual users. This study adheres to established ethical standards for social media data research, with careful attention given to privacy protection, data ownership, and responsible use. Given the psychological sensitivity of the inferiority theme, special care was taken to avoid stigmatizing language or interpretations.

4.2. Results

4.2.1. Semantic Boundary Identification

Figure 6 presents the geospatial-analogous semantic model based on the experimental data. The entire semantic space is divided into a series of non-overlapping polygonal regions, each corresponding to a specific semantic primitive. Compared with traditional semantic analysis models, such as the co-occurrence network model (Figure 7), the geospatial-analogous semantic model defines the semantic scope of each primitive through clear polygonal boundaries, resulting in distinct and recognizable semantic demarcation lines. The Chinese semantic primitives in the figure are translated into English in Appendix A.

As shown in Figure 6, the spatial structure of semantic regions displays clear heterogeneity, asymmetry, and hierarchical differentiation. Some areas have relatively regular, linear boundaries with compact internal structures and dense distributions, indicating high semantic correlation and coherence among the primitives within those regions. In contrast, other areas display irregular and curved boundaries, typically located near the intersections of different semantic themes, suggesting a certain degree of semantic blending, ambiguity, or gradual transition between adjacent primitives. Notably, in some transitional zones, the interlocking distribution of multiple polygons forms evident spatial gradients, reflecting the existence of semantic overlap or gradual semantic shifts.

Overall, the model clearly delineates the relative positions, boundaries, and transitional areas among semantic primitives, as well as the asymmetry within semantic structures, demonstrating strong representational capacity for semantic structure. Particularly in addressing the challenge of ambiguous semantic boundaries where traditional methods often struggle, the geospatial-analogous model offers a more intuitive, systematic, and interpretable form of visualization. These advantages strongly support its applicability and analytical potential in the structural modeling of textual semantics.

4.2.2. Semantic Clustering Pattern Identification

According to the results in

Table 2. Global semantic correlation diagnosis results.

	Value	Z-Score	p-Value	Variance	Expectation
Getis-Ord G	0.9994	0.3673	0.0478	0.0001	0.1992
Moran’s I	0.0304	2.3418	0.0120	0.0009	0.0027
Geary’s	0.5876	−1.2565	0.0410	0.0002	1.0000

, there is a certain degree of correlation among the semantic primitives. This finding validates the appropriateness of applying spatial analysis methods to semantic correlation research. Specifically, the Moran’s I result indicates a weak but statistically significant positive spatial correlation, reflecting the semantic correlation among adjacent semantic primitives. The Geary’s C result reveals differences between neighboring semantic primitives, suggesting semantic heterogeneity within the semantic space. Moreover, the Getis-Ord G results demonstrate a significant clustering of high-value semantic primitives (hotspots) in local regions, albeit with limited intensity but statistical significance. These results collectively confirm the existence of global semantic correlation within the semantic space, with each index capturing different aspects of global and local semantic spatial patterns. This, in turn, provides insights into the semantic spatial structure of the studied subject. The findings further validate the scientific legitimacy of employing spatial analysis in textual semantic research and offer critical support for understanding the semantic evolution dynamics of semantic primitives.

The local Getis-Ord G clustering map (Figure 8) reveals the spatial distribution patterns of semantic hotspots and coldspots. On the basis of confidence intervals, three significant levels emerge: red hotspot areas, concentrated in the left and upper-left regions, where related semantic primitives exhibit high-intensity clustering. Representative primitives such as 个性 (individuality), 博士 (PhD), and 婚礼 (wedding) reflect positive social attributes. In contrast, blue coldspot areas, occupying the lower-right and central regions, feature low-value clusters of primitives such as 本科毕业 (undergraduate graduation) and 悲观 (pessimism), which indicate negative emotional tendencies. The transitional zones display a random distribution without significant clustering patterns. This finding supports the hypothesis that semantic spaces exhibit polarity-based spatial differentiation [9,77,78].

The local Moran’s I analysis (Figure 9) further reveals the structural characteristics of spatial correlation. Four clustering patterns emerge in a gradient-like distribution: HH clusters (red) and LL clusters (blue) dominate the left and lower-right regions, respectively, forming a spatial opposition between positive semantic primitives (e.g., 博士 (PhD), 归属感 (sense of belonging)) and negative semantic primitives (e.g., 悲观 (pessimism), 理想 (ideal)). Notably, HL/LH clusters, which serve as transitional zones, are distributed along clustering boundaries, where primitives such as 单亲 (single parent) and 责任 (responsibility) have a semantic inhibitory effect on their surroundings. This core-periphery structure reveals the hierarchical relationships and propagation dynamics within the semantic network [8,58,72,78].

The local Geary’s C clustering map (Figure 10) presents a different perspective from local Moran’s I, omitting the low-high clustering (LH) category while incorporating a negative correlation type. In Figure 10, the HH clusters (red) are primarily concentrated in the upper-left and central regions, especially around semantic primitives such as 空虚 (emptiness), 印象 (impression), and 非议 (criticism). These areas exhibit high-value heterogeneity, reflecting core expressions and direct experiences related to inferiority, such as emotional fluctuations and negative sentiment articulation. Conversely, LL clusters are more widely distributed in the lower-right region, covering positive emotions and contextual variables rather than core drivers of self-abasement. Additionally, the Geary’s C classification results highlight boundary characteristics between regions. The transition zones between the HH and LL clusters exhibit significant heterogeneity, reflecting a semantic shift from core emotional drivers to contextual variables. These findings demonstrate that Geary’s C effectively captures local variations and boundary features in semantic space, providing a detailed analytical framework for understanding the complex spatial-semantic patterns of inferiority and their variable interactions [61,62,63].

4.2.3. Analysis of Knowledge Evolution

Figure 11, Figure 12 and Figure 13 present the 3D semantic landscape models constructed on the basis of the semantic clustering patterns shown in Figure 8, Figure 9 and Figure 10. In these figures, the height of each cube—resembling buildings in an urban landscape—represents the frequency of a semantic primitive within the corpus, collectively forming a semantic landscape composed of “semantic highlands” and “semantic lowlands”.

In Figure 11, Figure 12 and Figure 13, high-frequency semantic primitives such as 内心 (inner self), 容貌 (appearance), and 肌肤 (skin) are spatially concentrated and represented by taller cubes. These high-density regions exhibit compact spatial layouts, reflecting the central importance of the corresponding semantic topics within the studied domain. Other primitives, such as 爱情 (love), 地位 (status), and 外界 (external environment), although not the most frequent, still show significantly elevated cube heights relative to their surroundings, indicating secondary core positions in the overall semantic structure. These primitives are distributed across the middle-to-peripheral zones of the geospatial-analogous semantic space and are spatially separated from the primary core areas, illustrating a pattern of semantic diffusion from the core to the periphery. Additionally, terms such as 非议 (criticism), 兄弟姐妹 (siblings), and 好友 (friends), although located farther from the core, exhibit locally elevated cube heights, reflecting localized semantic prominence even in peripheral areas. This spatial distribution reveals distinctions in both the importance and relevance strength of semantic primitives within the studied context. Overall, the 3D semantic landscape model effectively visualizes differences in primitive frequency and delineates the spatial structure of semantic clustering, transitional boundaries, and peripheral zones. High-frequency primitives are easily identifiable through prominent cube heights, whereas semantic transition zones are characterized by gradual height changes and notable variations in spatial density, indicating semantic fuzziness and continuity at the boundaries. This structural representation provides a clear visual foundation for understanding the relative importance of semantic units, the spatial configuration of thematic domains, and the detection of localized semantic hotspots, thereby offering crucial technical support for knowledge evolution tracking.

Figure 14 illustrates the temporal dynamics of key semantic primitives across the years 2011, 2014, and 2017. The analysis focuses on high-value clusters identified by the local Getis-Ord G* index (HH), HL transition clusters, and HH clusters derived from the local Moran’s I and Geary’s C indices. The aim is to trace the evolution of expression intensity and clustering patterns over time. In Figure 14, semantic primitives such as 内心 (inner self), 容貌 (appearance), and 肌肤 (skin) consistently maintain hotspot status across all three time points. Notably, 内心 (inner self) registers the highest values throughout the study period and remains a stable HH cluster, indicating its sustained role as a core theme in inferiority-related discourse. 容貌 (appearance) and 肌肤 (skin) show similar trends and remain in close spatial proximity, suggesting a long-term semantic linkage between external appearance and internal self-perception. In contrast, the semantic primitive 爱情 (love) demonstrates a clear pattern of dynamic evolution. Initially located in a peripheral zone in 2011, it transitions to an HL cluster in 2014, expanding spatially, and by 2017, it becomes more prominent, forming localized co-clusters with 容貌 (appearance). This indicates a growing prominence of romantic concerns in the semantic space associated with inferiority, possibly reflecting the increasing emotional salience of intimate relationships. Another notable change is observed in the primitive 不漂亮 (not beautiful), which was not a hotspot in 2011 but emerged as an HL cluster in 2014 and became a prominent element in the negative sentiment sector by 2017. This trend suggests a rising prominence of negatively valenced appearance-related expressions in inferiority discourse. Additionally, although terms such as 外界 (external environment) and 地位 (status) do not match the frequency levels of core primitives such as 内心 (inner self), their spatial clustering has strengthened over time, indicating increasingly close co-occurrence patterns with central semantic themes.

5. Discussion

5.1. Overall

In this study, semantic boundary delineation was achieved by modeling the semantic space as a geospatial analog (Figure 6). This approach has several advantages over traditional textual semantic analysis methods. First, this model explicitly defines the boundaries of semantic primitives. The range and limits of each semantic region are clearly visible. Each polygonal edge represents a semantic boundary between two primitives. This is analogous to decision boundaries in nearest-neighbor classification algorithms and offers strong interpretability. In contrast, traditional methods like the co-occurrence network fail to present clear semantic boundaries. As shown in Figure 7, the relationships among semantic primitives tend to form intricate and entangled link structures in the co-occurrence network map. Moreover, the structural layout of the geospatial-analogous model enables effective identification of local semantic discontinuities, symmetry, and fuzzy transitional areas. Sharp and compact boundaries suggest abrupt semantic shifts. Fragmented, interleaving polygonal arrangements indicate gradual transitions or ambiguous overlaps between neighboring primitives. These subtle local variations are often difficult to visualize with conventional 2D methods. Additionally, the model provides a global structural view, allowing researchers to assess the size, shape, and adjacency of all semantic regions at the same time. This integrated perspective enhances the understanding of the overall semantic landscape and intertopic boundary dynamics.

This study presents a novel method for semantic correlation analysis. The method shows superior capabilities in detecting semantic clustering patterns, especially in terms of structural differentiation and multiscale adaptability. Traditional methods such as K-means or co-occurrence network analysis often assume homogeneous or static clustering structures. This makes them less effective at capturing the heterogeneity and transitional nature of semantic relationships [79,80,81]. In contrast, the proposed method employs three spatial measures: local Getis-Ord’s G*, Moran’s I, and Geary’s C. These measures allow for the detection of spatial associations and discrete patterns across different scales. This provides a more robust analytical foundation for semantic pattern identification [51,58,60,61,62]. Specifically, the Getis-Ord G* index identifies statistically significant clusters of high- or low-value semantic primitives. This reveals semantic “hotspots” and their spatial concentration patterns [59,60,82]. The Moran’s I index assesses global spatial correlation, highlighting polarity contrasts and stable cluster formations [51,58,59]. The Geary’s C index has heightened sensitivity to local variation. It detects subtle differences between neighboring semantic units, making it suitable for identifying boundary zones and transitional gradients [62,63]. The combined use of these measures enhances the sensitivity of this method to internal semantic variations. The proposed method is particularly effective for high-dimensional, nonlinear, and heterogeneous textual semantic data. It not only enables the detection of semantically characteristic regions, including core semantic zones and semantically symmetric regions, but also reveals weakly aggregated peripheral semantic units and their transitional connections to the central structure.

To represent knowledge evolution, this study presents a 3D method that maps the frequency of semantic primitives to spatial elevation. It simulates the structure of semantic space using polygonal partitions. Compared with traditional 3D representations (such as word clouds, heatmaps, co-occurrence networks, or topic flow charts), this method offers clear advantages. It expresses semantic prominence hierarchies, models structural distributions spatially, and captures temporal dynamics [79,81,83]. Conventional 2D methods primarily display static co-occurrence relationships. They struggle to convey the relative significance of semantic primitives or their trajectories over time. In contrast, the proposed method uses visual “height” to emphasize semantic units. This enhances the identification of primary, peripheral, symmetric, and asymmetric themes. It also improves spatial cognition within complex semantic structures. Moreover, this method is highly extensible and integrative. The use of ArcGIS-based tools supports both the quantitative validation of semantic structures and the temporal tracking of knowledge evolution [47,84,85,86]. By transforming changes in semantic frequency into spatial structural variations, the proposed method strengthens visual perception. It also improves the interpretability of the relationships among knowledge structures, semantic dynamics, and their evolutionary significance.

5.2. Significance

The significance of this study lies primarily in the innovative nature and practical applicability of the proposed framework, as well as its integration of interdisciplinary methods.

(1) Innovation. The proposed framework innovatively introduces both geographic metaphor theory and GIS methods into textual semantic analysis. Guided by geographic metaphor theory, it models the semantic space as a geospatial analog, which enables accurate delineation of boundaries among different semantic primitives [64,87]. By applying spatial measures to semantic correlation analysis, the framework facilitates the precise identification of semantic clustering patterns. Furthermore, it intuitively represents knowledge evolution through a 3D semantic landscape via ArcGIS-based visualization techniques [88,89,90]. These innovations significantly enhance the interpretability and visualization capability of textual semantic analysis.

(2) Practical application value. The experiments on real-world social media data with inferiority themes demonstrate the efficacy of the proposed framework in handling large-scale, unstructured, high-noise, multisource, and heterogeneous datasets. Thus, the framework has the potential to be applied to practical applications such as public sentiment analysis, collective psychological monitoring, and trending-topic tracking. Moreover, the proposed method for representing knowledge evolution provides valuable analytical support for practical decision-making, risk assessment, early warning systems, and policy formulation.

(3) Interdisciplinary methodological integration. This study realizes an effective synthesis of geographic information science, cognitive linguistics, and natural language processing. Geographic metaphor, as an interdisciplinary theoretical perspective, creates analogical mappings between abstract semantic concepts and spatial concepts from geography [64,87]. This interdisciplinary integration not only enriches the theoretical perspectives traditionally employed in semantic studies but also opens new methodological paths for applying GIS methods to textual semantic analysis [85,88,91,92].

5.3. Limitations and Future Work

Despite the aforementioned contributions, this study has several limitations with respect to methodology, experimental data, and results:

(1) Methodology. While this study demonstrates both the practicality and validity of modeling semantic space as a geospatial analog, it is worth noting that semantic space and geographic space are not ontologically equivalent. The spatial structures in the semantic space are primarily metaphorical constructions. These are derived from the dimensionality-reduced embedding of semantic vectors [17,18,19,66,79]. Therefore, interpretations of semantic clustering and evolutionary patterns should be contextually bounded. They should also be informed by linguistic and semantic knowledge to avoid misinterpretation driven by spatial intuitions [20,51]. Although the proposed framework is grounded in the assumption that spatial proximity indicates semantic correlation, several inherent theoretical and methodological limitations should be acknowledged. For example, t-SNE excels at preserving local structures, but its global topology may not fully reflect the original high-dimensional semantic space [50,93,94]. In addition, spatial statistical methods are fundamentally based on Euclidean distance. This limits their ability to capture nongeometric semantic factors, such as pragmatic functions or polysemy. Future work could consider the use of topology-preserving algorithms (e.g., UMAP and GNN) and expert-driven semantic evaluations to further assess the validity and robustness of this spatial mapping strategy.

(2) Data. The experimental data focus exclusively on Chinese social media texts related to the theme of “inferiority”. The informal, unstructured, and heterogeneous nature of social media language introduces valuable complexity. This makes the dataset a challenging yet effective testbed for evaluating the proposed framework [46,95,96,97,98]. However, the data also have inherent limitations [68,71,72,98,99,100]. First, the user base of Sina Weibo may introduce selection biases, since it does not represent the full demographic spectrum of the Chinese population. Second, regional linguistic variations and culturally specific expressions may affect the interpretation of semantic patterns. Third, the high level of noise and nonstandard language usage may compromise the accuracy of semantic clustering and inference. Future efforts should focus on evaluating the framework across multilingual and cross-platform datasets. This would help to assess its cultural adaptability and semantic stability under diverse linguistic and contextual conditions.

(3) Results: This study empirically validates the effectiveness of the proposed framework with respect to semantic boundary delineation, semantic pattern recognition, and knowledge evolution representation. However, the absence of expert validation from relevant domains has limited the interpretation of results. This may allow certain deficiencies of the framework to go unnoticed. Future work will focus on improving the interpretability and applicability of the model. This can be achieved by introducing adaptive modeling techniques, integrating auxiliary data, and enhancing validation strategies at the cognitive level.

6. Conclusions

This study proposes an innovative textual semantic analysis framework that integrates geographic metaphor theory and GIS methods. The efficacy of the framework in semantic boundary delineation, semantic clustering pattern identification, and knowledge evolution representation has been validated on real-world social media data.

With respect to semantic boundary delineation, the proposed framework models the semantic space as a geospatial analog, which not only facilitates accurate delineation of semantic boundaries but also enables precise identification of core, peripheral, and symmetric semantic regions. In terms of semantic clustering pattern identification, the framework enables precise detection of global semantic correlations and local spatial semantic clustering patterns (e.g., HH, HL, LH, and LL clusters) within the semantic space, while also allowing for intuitive visualization of these patterns and their features—such as semantic symmetry—through two-dimensional visualization techniques. The presented analyses reveal distinctive features and roles of these clustering patterns within semantic space. Specifically, HH-type clusters consist primarily of highly abstract and symbolic semantic primitives; HL and LL clusters predominantly contain concrete concepts or terminologies with lower abstraction levels, exhibiting inhibitory effects on neighboring semantic primitives. With respect to knowledge evolution representation, the framework visualizes knowledge evolution in a form analogous to urban development and expansion, integrating semantic structures and evolutionary processes into a coherent and intuitive visual representation.

Despite these contributions, certain limitations remain. For example, the high noise and non-standardized characteristics of experimental data might reduce the accuracy of semantic clustering identification. In addition, although the t-SNE algorithm excels at preserving local neighborhood relationships, it may distort the overall positional relationships among semantic primitives, failing to retain global structural integrity. Finally, the interpretation of semantic clustering patterns in this study lacks theoretical support from linguistics. This shortcoming may lead to insufficient identification and interpretation of semantic patterns.

Future work will focus on addressing the above limitations. First, multilingual and cross-platform datasets will be employed to increase data compatibility. Second, coupling the t-SNE algorithm with topology-preserving algorithms such as UMAP and GNN will ensure the retention of global structural integrity. Finally, interdisciplinary collaboration, particularly with domain experts, will be fostered to conduct more comprehensive and specialized analyses of semantic patterns.