Dependency Grammar

Dependency grammar organizes relations among linguistic units, such as words, according to their predominant or dependent roles to communicate meaning (fig. 1). The verb is taken to be the structural center of a clause and requires the semantic agreement of other words directly or indirectly connected to it according to its dependency direction, as in “is dependent on” or “has dependent.” The data structure is represented by the dependency arcs between the main word acting as the head and the head’s dependents. Dependency structures may be formed as chunks consisting of nouns with auxiliary relations to verbs, such as modifiers or other information, adding complexity, such as phrases.

Part-of-Speech Classes

SpaCy natural language processing pipelines tokenize phrases and assign POS abbreviation tags to those tokens and tags for the dependency relations determined among tokens based on Universal Dependencies libraries, as presented by Marneffe and others (2024). Token texts and their associated tags compose the tabular data column headings of the USGS data release associated with this report (Abbott, 2024). The text analysis in this study produced eight POS classes of content words for object representations and four classes of function words for operational applications; these are listed below. Function words provide clarifying context to descriptions. For example, descriptions formed with the determiner “the” that appear to refer to exactly one object could indicate logical primitives.

The eight POS content words and their abbreviations are:

The four POS function words are:

A token's lemma is an attribute that describes the root form or infinitive of that word. In the word’s dictionary form, a lemma is used as a heading indicating the subject or argument of an annotation or a dictionary entry. Lemmas may be preferred forms for entity classes in an ontology because definitions usually reflect stable states such as roots or infinitives. However, linguistic class definitions include conditional properties indicated by the word token texts of their input clause. Word token texts vary to indicate semantic properties (such as person, number, tense, use, and the action role of subjects and objects) that may be transferable to graphs. POS tags, such as “noun,” “verb,” “adjective,” and “adverb,” indicate grammatical roles that help distinguish how a lemma may function within a graph, whether as a class, relationship, or property.

The POS sets transfer to the class-subclass hierarchy of taxonomy and tree-type databases to allow Boolean operations involving set members and inference. However, the relations and attributes may lack equivalent formal semantic axioms or formulas that enable some semantic similarity measures but limit logic applications. Some set relations are inherent in definitions of words, such as plurals. Different kinds of relations among objects, space, and time, such as agency, possession, or topology, are marked by prepositions or subordinate conjunctions. Verb relations and attributes can infer some processes. For example, in the meteorological phenomenon called the convection cycle, dynamic markers such as “cooling,” “shrinking,” and “sinking” can be used to define an entity such as a convective storm, “An atmospheric storm which is generated by the heating and convection of moist and unstable air masses” (The Ontology Lookup Service, 2025). Adverbs can reflect conditions of necessity, uncertainty, possibility, or permission, including time conditions. Some subordinate conjunctions beginning with “wh-,” such as “where” and “when,” provide indications of space and time.

Grouping POS with tags categorizes syntactic text terms into semantic roles and creates a lexicon. The lexicon offers quantitative summaries of the input data corpus. Each word has its lemma listed once in order of most frequent occurrence, with POS category tags included for grouping terms. Descriptive parts of speech, such as determiners, are removed from the lexicon. The same format is applied to chunks. Determiners are like noun modifiers when included in chunks, which are groupings of a noun and its modifiers. Chunking streamlines the processing workload if context is not a priority.

Dependency Relation Properties

Dependency relations are grammatically associated with certain parts of speech to form semantic patterns. The word token text has attributes of its POS and its lemma. The head, representing the predominant concept of the relationship, has attributes of the POS and the lemma. The dependency relation has one head and any number of children.

The dependency relation for the head and its children is similar to an ontology property’s domain and range class. The following relations form the semantic transformation rules for the grammar ontology:

Table 1 shows an input clause, “An aircraft facility is an area where aircraft can take-off and land usually equipped with associated buildings and facilities,” tokenized as words or as chunks and assigned their corresponding POS categories, and the dependency relation of the tokens to their arc heads of phrases within the clause.

An input clause can have multiple dependency relations that can form subgraphs within the same input sentence or across other definitions. Dependent children and their relation to the heads form multiple arcs within an input sentence to map to triples.

To simulate the node-edge-node pattern of triples, the dependency relations between words grouped by the POS of the data output by the natural language processing (NLP) software were directed edges, which connect the child to the head in the direction from child to head. However, directed edges can appear anywhere in the clause, so the meaning may be represented as pointing right or left. The object dependency represented by the triple property is assumed to be roughly the same as in grammar in that the object is determined by, and secondary to, the subject. However, the triple syntax can be specified using an inverse property relation, creating a directed edge to the left or right of the specified nodes. Care is required to confirm directionality when mapping the dependency of the translation because some linguistic semantic dependencies are similar but not equivalent to the formal logic.

Some dependency relations between words of an input definition are related to a baseline triple based on a verb, its subject, and any object. Adding additional properties to these main parts of a triple creates chunks, which are nouns and their modifiers (usually adjectives). Chunks can be mapped to subgraphs in the positions of the subject and object nodes.

To produce the data for the topographic subject area study, NLP preprocessing was completed using two approaches. One approach was to analyze each input token to create chunks for the head of the dependency arc. The chunk pattern approach combines related dependent tokens to group commonly used POS, such as articles and determiners, or to document words commonly used together. The second approach, the token pattern approach, individualizes each unit by category to be examined by the category’s clausal roles. Chunks are phrases formed as the head, dependency, and children of the subject and object relations of the verb of an input phrase. Modifiers and other dependents form clausal subjects with the head of the chunk, the core of the subject, or the object of the basic triple. In table 2, the arc heads “is,” “in,” and “of” are word tokens. “Any bowl-shaped depression” and “the surface” are chunks. The differences between these two approaches to information retrieval are the number and pattern of categories and properties.

Part-of-Speech Rules

The number of POS terms output from the NLP was counted (table 3). Nouns predominate but include modifiers essential to the meaning of the nominal (the combined modifiers and noun). Other adjectives are descriptive. In addition to over 200 verbs, all other POS terms number fewer than 100, including modifiers and function words, which include prepositions that combine with nouns to suggest spatial relations and conjunctions that join more than one phrase that may indicate context.

Parts of speech other than nouns and verbs use dependency relations that add semantic specificity to the main parts of the triple through the data output from tokenization. Dependency relations analyses break up the input clause to be readable by logical relations (table 4). In this way, the data can be reused to connect phrases to each other in ways that support graph processing.

The twelve types of POSs are grouped into four roles (fig. 3). The chunk is based on the nominal consisting of a noun, adjectives, and other token dependencies of the noun. The property role includes the verb and adverb. The third position, object, includes the adposition or preposition and directly follows the property; the object has a chunk pattern. The function role of conjunctions is to join or provide context for chunks. The semantic schema mapping rules to organize the chunk, property, and function are explained in the section of this report titled “Semantic Schema Mapping.”

Semantic Schema Mapping

The semantic schema maps grammar rules to generate types of triple patterns, particularly RDF. The property establishes the primary semantic element with subject and object arguments. The subject is the head of the clause. The subject and property without an object fulfill the requirements of a clause. As a triple, the object can be implied, such as with a blank node. The subject and property have an object, and objects and subjects may have modifiers. Function words join or describe a context for multiple subjects and objects.

The dependency relations used as properties for triples have domain and range classes corresponding to grammar classes within the ontology structure. The SVO triple pattern has five subtypes of POS classes for the lexical graph data that will be queried using SPARQL (fig. 4).

Complexes of noun modifiers have four potential property patterns, as do verbs and their modifiers, to support property clusters. An object-like pattern called a chunk is used as the subject and reused as the object and as part of combined triples (fig. 5). Chunks have nominals such as modifiers and subtypes.

Parts of speech relating to the central grammatical verb combine to form the action property (fig. 6).

Retrieval variables for a data sample forming a verb property and its modifiers include adverbs (ADV), auxiliary (AUX) verbs, and the central verb (VERB). Each subproperty of the dependency relation property controls the appropriate structure and function relations of POS word instances to be represented in RDF.

Directionality

An object’s dependency on a head preserves semantic coherence as a graph among tokens or POS categories to resemble the grammar of phrases and complex concepts. The head is always the subject, and the dependency direction is from children to head. The directionality of the RDF is in the inverse direction, from subject to object. A triple is read from left to right, although dependents can appear on either the left or right side in a dependency relation.

Data were formatted as a dependency arc starting at the head, shown in table 5 as HeadText, the dependency relation abbreviation, and a token representing the child. The children are more variable than the subject, so they are not grouped as often in a POS category, or the category is left blank (table 5). The arc head is analogous to the triple subject, and the token dependent aligns with the object as the argument of the verb.

Force Dynamics

The topographic analysis of this study focuses on property patterns, including geographic process and force dynamic concepts, which include morphology and description, properties of human intention, and concepts of spatial data science. The property and lexical semantics of feature-type definitions build knowledge more flexibly than a literal clause by relating the data as triple propositions for graph processing. For example, a definition from the database for “arête,” a landform type, would be “a sharp narrow ridge formed as a result of glacial erosion from both sides.”

Changing the representation as an SPO pattern, in which case terms or labels take the three positions of a potential RDF serialized triple, creates the following series of ontology propositions to form the criteria of inclusion for a set of objects and attributes that comprise a dataset schema for information retrieval.

The style of writing reflects the normalization of the input data adapted for the ontology taxonomy. Spaces between complex terms are omitted to indicate a single-class label. Object classes with multiple members are written in single numbers.

The propositions listed above build a pattern that fits a subgroup of the dataset, defining natural features of a morphological process, which entails the determination of a shape that is described verbally (like a metaphor) or geometrically. The shape helps define the feature-type class along with supporting details. The dependency of subjects, properties, and objects in the pattern connects the feature type to additional data such as its identity or context.

Social Space

The language analysis in this study inferred human intention and the agency of socially constructed properties. For example, a park is a place or area that, set aside by human agents, will allow for specific attributes while suppressing other competing attributes. Socially constructed spatial relations, such as “set aside” or “preserve,” may seem more like environmental attributes and not specific spatial relations with physical manifestations. The distance and implicit nature of human influence control the environment surrounding or setting conditions within which an entity operates. Spatial relations are required for entities to exist and operate in their environments. Socially constructed spatial relations can be understood as informal logic relations between sets of entities.

Cultural geographers have analyzed socially constructed spatial relations and found that they were not adequately modeled in Geographic Information Systems (GIS). Massey (2004) found that humans create and control spaces of interrelations. In those relations, temporal change differs from natural change, and socially constructed relations involve quicker observable temporal change in general. The outcomes of morphologic processes are less determinant and more varied than those of long-lasting natural processes.

The G2G approach of listing natural language propositions to align with semantic technology standards, as was shown for an arête in the “Force Dynamics” section of this report, was applied to a group of definitions involving socially constructed spatial relations, also known as cultural features. The key aspects of cultural-features morphology were agents, objectives, and technology that broaden the scope of topographical shapes. An example of statements defining cropland as “land that is plowed or cultivated for crop production” appears below as a list of necessary classes and properties forming the criteria of inclusion.

As demonstrated by the semantic analysis above, topographic morphology is driven by agents with objectives using tools, based on the natural environment. Cropland varies in shape based on this pattern of interaction. The shape of the cropland area is determined by the interactions as, for example, mixed-use cultivation such as a vineyard designed for small-scale production on a larger farm, cropland driven by circular irrigation systems, or cultivated terraces adjusted to contoured terrain.

The semantic patterns of cultural features were initially similar to the natural feature patterns, but some differences were evident. Morphologic processes included physical force dynamics and human-driven, tool-based forces. The “shape” descriptor, important for natural features, is replaced by a term of human interaction, such as “cultivated.” In the natural feature example in the “Force Dynamics” section of this report, all arêtes would look similar based on their definition and generalized geometric shape. However, this is not so for all croplands. Socio-spatial relations are affected by environmental conditions and the cumulative history of human modifications; the abandonment of cropland to physical forces also affects relations. These relations are often observable in the topography of places and regions.

Subgraphs of Natural and Cultural Feature Properties

A use case to guide queries along semantically specific property paths differentiates natural and cultural features using force dynamics and social space patterns, beginning with a general morphology class whose criteria are inherited by both natural and cultural features. The use case concept relies on common understandings about topography for testing this graph-based querying approach and is easily altered for other general-purpose uses.

Two knowledge concepts to disambiguate features using semantic graph searches are commonality, for which we heuristically assume a natural feature is the entailment or predecessor of cultural features, and cause and action. Shown in figure 7, commonality is tested if there is a missing edge between object classes or if there are duplicate nodes. The cause and action pattern tests for a common ancestor node or alternative action pathway along edges.

Commonality is tested if an edge between object classes is missing or if there are duplicate nodes. The cause and action test is done for a common ancestor node or an alternative action pathway along edges. Three definitions from the glossary database for the feature types “bridge” and “arch” that pose a logical contradiction are used for semantically differentiating the data for this use case:

The tokenized definitions of arch and bridge were mapped for commonality and cause and action patterns using the natural or cultural feature properties (fig. 8). The feature-type patterns shown in blue in figure 8 are the same pattern as in figure 7, with the definitions for arch and bridge in black print (fig. 8). The semantic paths of term definitions are linked by properties. Topological relations to geospatial features are included in figure 8.

Commonality is tested by missing edges or duplicate nodes in the subgraph of the input clause token nodes. “Bridge” includes a property indicating a purpose, which is missing for “arch.” The commonality (fig. 8A) of the two feature type classes is changed by “carry,” which disambiguates the cultural feature-type bridge from a natural bridge.

Differences in cause and effect are matched against the subgraphs. “Cause and action” is defined as “common ancestor or succession” on nodes and “alternative action” on edges. Arch and bridge have a common ancestor, Agent, and can have a common succession to the class Shape, which is shown in figure 8B by blue arrows, and as an alternative action differentiating artificial and natural semantic paths, which are shown by purple arrows.

The semantics of shape may explain that a bridge is the same thing as an arch if a bridge includes an overhead linear connection from point to point. Roughly speaking, bridge and arch share the shape of an overhead point-to-point stationary connection. A bridge with the same natural features as an arch, without a purpose, could be the same thing as an arch. When the purpose “carrying across” is added, the bridge shape differs from an arch.

The use case demonstrates how graph properties can support semantic specification to resolve potential issues of ambiguity, differentiation, and relative meaning. Graph data processing depends on patterns for these functions. The “Results” section clarifies the property patterns for some topographic semantic patterns.