Natural language is inherently multi-dimensional. When humans process an utterance, we do not just see a flat sequence of characters or words. We simultaneously decode its morphological structure, its surface syntax, its deep semantic relationships, and its broader pragmatic or cultural context.
Traditional computational approaches often collapse this rich hierarchy into a single rigid structure or a flat sequence of tokens. This flattening forces a destructive trade-off: optimize for surface-level processing (like tokenization) or high-level understanding (like intent), but lose the explicit links between them.
Multi-layered linguistic representation solves this by treating text processing as an additive, stratified process. Instead of replacing or altering the original text, systems stack independent but synchronized layers of linguistic information on top of it. As Yoshihiko Hayashi notes in A Linguistic Service Ontology for Language Infrastructures, this additive strategy ensures that the raw text remains a stable foundation while enabling increasingly abstract levels of analysis to coexist.
The Stratified Pipeline: From Surface to Deep Syntax
To understand why this works, consider how language can be broken down into progressive tiers of abstraction. In machine translation, for example, Gorka Labaka et al. describe the TectoMT architecture in Deep-syntax TectoMT for English-Spanish MT, which divides language into four distinct layers:
- w-layer: The raw, physical text string.
- m-layer: The morphological tier handling word forms and categories.
- a-layer: The shallow-syntactic structure mapping surface dependencies.
- t-layer: The deep-syntactic or "tectogrammatical" tier capturing underlying semantic roles.
By separating these concerns, a system can resolve ambiguities step-by-step. A morphological layer handles word forms, while a deep-syntax layer handles abstract semantic roles regardless of how they are physically arranged in a sentence.
This stratificational parallel processing, formalized by František Mráz et al. in Towards a formal model of natural language description based on restarting automata with parallel DR-structures, allows a single synchronized framework to map surface-level features (like prepositions) directly to deep semantic attributes. This mirrors early connectionist parsing insights from Steven L. Small et al. in Toward Connectionist Parsing, who structured linguistic data into a hierarchy of lexical, word-sense, and case-logic levels to transition smoothly from syntax to semantics.
[Deep Syntax / t-layer] ---> Abstract Semantic Roles (e.g., Actor, Patient)
|
[Shallow Syntax / a-layer] ---> Surface Dependency Trees (Subject, Object)
|
[Morphology / m-layer] ---> Lemmas, Parts of Speech, Inflections
|
[Raw Text / w-layer] ---> "The chef cooked the meal."
Standoff Annotation and Graph Architectures
If we have multiple layers of annotation, how do we store and access them without creating a tangled mess of inline tags? The answer lies in standoff annotation and hierarchical graph structures.
Instead of embedding annotations directly within the text (which fails when annotations overlap or are discontinuous), modern frameworks decouple the metadata from the source. In the MEANING Annotation Format (MAF) introduced by Emanuele Pianta et al. in Representing and Accessing Multilevel Linguistic Annotation using the MEANING Format, higher-level semantic or syntactic units do not point to raw text strings; instead, they point to lower-level morphological tokens. This creates a stable, object-oriented hierarchy.
Similarly, Ekaterina Buyko et al. utilize inheritance within an annotation type system in An Annotation Type System for a Data-Driven NLP Pipeline to organize data into document structure, syntax, and semantics, allowing domain-specific features to be added without breaking pipeline compatibility.
When scaled to entire language resources, this can take the form of a hierarchical graph. For instance, the BILGraph framework developed by Fabio Cozman et al. in Unified Knowledge-Graph for Brazilian Indigenous Languages: An Educational Applications Perspective organizes data into four distinct layers—sentences, tokens, lemmas, and high-level semantic concepts—creating a multi-tiered graph that preserves granular token details while linking them to abstract meanings that transcend specific language instances.
Cross-Layer Synchronization and Querying
The true power of a multi-layered approach is not just storing the data, but querying across the layers simultaneously. To understand why this works, consider the requirements of complex linguistic search.
In the Prague Dependency Treebank 2.0, analyzed by Jirí Mírovský in PDT 2.0 Requirements on a Query Language, effective search requires the ability to synchronize queries across multiple layers. A researcher might want to find every instance where a deep-syntactic semantic role (e.g., the "actor") is realized by a specific surface-level word position or grammatical case. Dynamic reference binding across layers allows the system to map abstract semantic functions directly to physical sentence structures, revealing the diverse ways meaning manifests on the surface.
Handling Real-World Complexity
Multi-layered representation is particularly effective for linguistic phenomena that defy flat, single-tier analysis, such as code-switching, idiomatic expressions, and pragmatics.
Code-Switching and Dialects
When speakers mix languages or dialects, a flat model struggles to capture both the syntactic mechanics and the overall intent. Mona T. Diab et al. in Creating a Large Multi-Layered Representational Repository of Linguistic Code Switched Arabic Data and Aditya Bohra et al. in A Dataset for Detecting Irony in Hindi-English Code-Mixed Social Media Text demonstrate that by using multi-level annotations—tagging individual tokens with specific language or switch-point markers while assigning global semantic labels to the entire sequence—models can study how fine-grained syntactic transitions influence overall sentiment or irony.
Cultural and Idiomatic Nuance
Literal translations often miss the point of cultural expressions. To benchmark large language models on Arabic proverbs, Samar Mohamed Magdy et al. in JAWAHER: A Multidialectal Dataset of Arabic Proverbs for LLM Benchmarking structured data with distinct layers: literal translations, idiomatic equivalents, and detailed cultural explanations. This multi-layered semantic annotation makes it possible to evaluate whether a model understands the surface text, the underlying meaning, or the historical context.
Pragmatics and Discourse
In analyzing spoken language, flat transcriptions omit how social context alters speech. In Toward mapping pragmatic impairment of autism spectrum disorder individuals through the development of a corpus of spoken Japanese, Sumi Kato et al. organized features into ideational (describing reality) and interpersonal (managing social relationships) metafunctions. This allows researchers to isolate how specific linguistic choices—like negotiating particles—map directly to neurocognitive functions or social communication patterns.
Key Foundational Work
A Heuristic Approach to Natural Language Processing (1965) — contributed 2 concept elements to this direction.
Syntaxe Et Interpretation (1965) — contributed 1 concept elements to this direction.
The Application of FORTRAN to Automatic Translation (1967) — contributed 1 concept elements to this direction.