When we translate a sentence from one language to another, we rarely map words directly. Instead, we read the source text, extract its core meaning, and reconstruct that meaning using the grammar and vocabulary of the target language.
In computational linguistics, this intermediate, language-free meaning layer is known as an interlingua or a language-independent semantic representation. The core intuition behind this approach is that language-specific syntax, morphology, and word order can be treated as a surface-level "style" or wrapper. By stripping away this linguistic variance, we can map sentences into a neutral, intermediate space that captures pure conceptual meaning.
Language A (Surface Syntax) ───► [ Language-Independent ] ───► Language B (Surface Syntax)
Language C (Surface Syntax) ───► [ Semantic Space ] ───► Language D (Surface Syntax)
This architecture provides a massive scaling advantage. Instead of building separate transfer rules for every possible language pair—an $N \times M$ bottleneck—systems only need to map inputs into the shared semantic space and generate outputs from it.
The Symbolic Foundation: Conceptual Networks
Early computational frameworks built this language-free space using explicit symbolic networks. Rather than relying on syntactic trees tied to a specific language's grammar, researchers sought to map natural language directly to real-world concepts, actions, and entities.
This paradigm was pioneered by Schank et al. (1969), who introduced Conceptual Dependency (CD) theory. They argued that sentences should be parsed into a language-independent network where nodes represent universal word-senses, actors, or locations, and edges represent meaning-contingent dependency links (such as actor-action-object relations). Because this representation prioritizes semantic roles over surface syntax, systems like MARGIE (Schank et al., 1973) and SAM (Cullingford, 1977) could take an English sentence, map it to a pure conceptual form, and then re-translate that core concept back into numerous semantically equivalent paraphrases or entirely different tongues.
In modern natural language processing, this graph-based philosophy survives in formalisms like Abstract Meaning Representation (AMR). As outlined by Schneider et al. (2015), AMR abstracts away from surface syntax to capture the "who did what to whom" of a sentence. By treating translation, summarization, and parsing as transformations within a common semantic graph space, researchers can bypass direct string-to-string mappings entirely.
Resolving Structural Divergences
A primary reason language-independent representations work is their ability to elegantly resolve structural and thematic divergences between languages. Different cultures and linguistic families express identical ideas using fundamentally incompatible grammatical strategies.
For example, Dorr (1990) focused on "thematic divergences," where languages swap subjects and objects or promote modifiers to main verbs to convey the same meaning (e.g., the structural difference between "I like grapes" and the Spanish "Me gustan las uvas"). By using a language-independent interlingua called Lexical Conceptual Structure (LCS), the system decouples core meaning from language-specific syntax. The source syntax is mapped to a universal conceptual form, and the target syntax is generated cleanly from that shared abstraction, eliminating the need for messy, language-pair-specific transfer rules.
This decoupling applies across various linguistic complexities:
- Indirect Speech and Intent: Zajac (1989) showed that complex, indirect speech acts and honorifics can be resolved into abstract communicative intents, allowing for natural translations between structurally disparate languages like Japanese and English.
- Quantification Mismatches: Guan et al. (2011) demonstrated that mapping nominal classifiers (measure words) to a common semantic core allows machine translation systems to bridge structural differences in how languages handle quantification.
- Spatial Relations: Munro et al. (1990) utilized a language-independent spatial semantic representation to capture geometric and relational features, successfully bridging the conceptual gap between context-dependent locative prepositions (like "on," "at," and "in") across languages.
The Neural Emergence: Latent Interlinguas
With the advent of deep learning, the field shifted from hand-crafted symbolic graphs to continuous mathematical spaces. Instead of human experts defining the conceptual primitives, neural networks are forced to discover them.
Tiedemann (2018) demonstrated this phenomenon by training a single neural machine translation (NMT) model on massive multilingual corpora. By using translations as "semantic mirrors," the model naturally organizes different languages into a shared geometric space. Remarkably, this mathematical space clusters languages in a way that mirrors real-world linguistic families and genetic relationships, entirely without prior geographical or historical knowledge.
To explicitly exploit this behavior for zero-shot cross-lingual transfer, Hong et al. (2020) developed an unsupervised adversarial framework that maps sentence embeddings into a neutral, intermediate mathematical space. By penalizing the model when it can identify which language an embedding originated from, the system effectively strips away linguistic variance. This isolates the underlying semantic content, allowing a model trained only on English task data to generalize to low-resource languages without any parallel training text.
Language A Embedding ───┐
▼
Language B Embedding ───► [ Adversarial Filter ] ───► Pure Semantic Vector (Language-Agnostic)
▲
Language C Embedding ───┘
Proving the Existence of Language-Free Concepts
Do large language models (LLMs) truly understand concepts separate from words, or are they merely performing sophisticated statistical token matching?
To answer this, Dumas et al. (2025) utilized activation patching to inspect the internal layers of modern transformers. They discovered that transformers develop internal representations of concepts that are entirely independent of the specific language used to express them. By isolating these hidden states and averaging a concept's representation across multiple languages, the authors found they could actually improve the model's downstream translation performance.
This empirical evidence confirms what early computational linguists hypothesized: deep language models naturally converge on a universal semantic space, successfully separating tongue from thought to achieve true cross-lingual understanding.