Human language is fluid, ambiguous, and deeply dependent on context. A single sentence can convey subtle nuances through metaphor, syntax, or tone. In contrast, a knowledge graph (KG) is rigid, discrete, and explicit. It organizes information into strict relational triples—such as (Diabetes, Type_Of, Disease)—where entities and relations have fixed, unambiguous definitions.
When building intelligent systems, we often want to combine the flexible reasoning of natural language models with the factual precision of knowledge graphs. However, doing so introduces a fundamental modality gap.
Because text and graphs are structured differently, neural models traditionally project them into isolated, non-overlapping vector spaces. An embedding for the word "diabetes" generated by a language model shares no geometric relationship with the embedding for the node "Diabetes" generated by a graph neural network. This representation gap makes it incredibly difficult for models to effectively fuse structural facts with linguistic context.
To bridge this gap, researchers have developed three primary strategies: aligning vector spaces through joint pre-training, designing mutual attention mechanisms for cross-modal interaction, and linearizing graphs to process them directly within language models.
The Root of the Mismatch: Isolated Semantic Spaces
To understand why combining text and graphs is difficult, consider how each modality learns representations. Language models learn by predicting words based on sequential co-occurrence. Graph embeddings learn by preserving topological structures and relational distances.
As Kartsaklis et al. (2018) point out, this creates a structural mismatch between sequential, fluid text and hierarchical, rigid graph representations. This mismatch prevents simple alignment between descriptive phrases and their corresponding entities. Furthermore, Christopoulou et al. (2021) observe that text is context-rich and unstructured, while graph facts are context-free and structured.
When a model attempts to process both modalities simultaneously—for example, concatenating a text embedding with a graph embedding—it fails because the two vectors speak different mathematical languages. As Ye et al. (2023) demonstrate, this distributional gap means the model cannot easily reconcile linguistic nuances with explicit factual relationships, leading to suboptimal information fusion.
Strategy 1: Joint Pre-training and Shared Anchors
The most direct way to bridge the modality gap is to force text and graph encoders into a single, unified semantic space during training. This is achieved by using "anchors"—entities that exist in both the text (as a word or phrase) and the graph (as a discrete node).
Unstructured Text: "The patient was diagnosed with diabetes..."
|
[Shared Anchor]
|
Knowledge Graph: (Diabetes) --[Type_Of]--> (Disease)
In the JAKET framework introduced by Yu et al. (2022), the model uses textual descriptions to initialize graph nodes, while simultaneously using the graph's structural topology to enrich the language model's text representations. By allowing the language model and the graph components to "talk" to each other during pre-training, the entities are bound to the same semantic coordinates.
Similarly, Sakhovskiy et al. (2025) developed the BALI framework, which synchronizes the internal logic of medical knowledge graphs with the linguistic patterns of language models. By leveraging shared entity anchors during training, the model learns to map linguistic mentions directly to their structural neighbors without requiring extra computational overhead during inference.
Instead of just aligning nodes, Dou et al. (2020) proposed Conceptual-Contextual (CC) embeddings. This approach injects structured knowledge by training a contextual encoder to satisfy graph relational constraints (such as the translation principle $h + r \approx t$) directly within real-world textual mentions. This forces the model to create "concept-aware" text representations that inherently understand graph topology.
Strategy 2: Mutual Attention and Cross-Modal Alignment
Rather than forcing static embeddings into the same space, another approach dynamically bridges the gap during inference using specialized attention mechanisms. This allows the text to guide how the graph is read, and vice versa.
Han et al. (2018) introduced a mutual attention mechanism that exemplifies this bidirectional synergy. In their framework, the structured knowledge graph acts as a filter to reduce noise and ambiguity in unstructured text. Simultaneously, the rich semantics of the text are used to refine and update the structural representations of the graph.
[Text Sequence] <--- Mutual Attention ---> [Graph Triplets]
(Filters noise) (Provides structure)
To make this alignment more robust, the Graph Reasoning Transformer (GRT), proposed by Zhao et al. (2024), shifts the focus from individual nodes to entire triplets. By treating a triplet (head, relation, tail) as an atomic unit of information, the model uses a cross-modal alignment pre-training objective to map linguistic tokens directly to structured facts, bypassing the limitations of simple embedding concatenation.
This multi-angle view is echoed by Deng et al. (2019), who utilize a multi-view attention mechanism to bridge the gap for question answering, allowing unstructured questions and structured database facts to mutually reinforce each other's representations.
Strategy 3: Linearization and Generative Translation
A third, increasingly popular intuition bypasses the vector space mismatch entirely by translating the graph's structure into a textual format. If a graph can be converted into a sequence of words, a standard language model can process it without needing a separate graph encoder.
In the ReGen framework by Dognin et al. (2021), graphs are linearized into sequences (e.g., translating a set of edges into a standardized string format). This allows pretrained language models like T5 to perform bidirectional translation—moving seamlessly from Text-to-Graph (T2G) and Graph-to-Text (G2T) while maintaining semantic accuracy.
Graph Structure: Linearized String:
(A) --[born_in]--> (B) =======> "A | born_in | B"
Alternatively, language models can be used to generate graph-like queries from text. Pham et al. (2025) developed ClaimPKG, which uses a lightweight, specialized LLM to translate unstructured text claims into "pseudo-subgraphs." These subgraphs serve as precise search queries that bridge the modality mismatch, allowing fact-checking systems to retrieve exact matches from formal graph databases.
Why It Matters: Real-World Applications
Successfully bridging the semantic-structural gap unlocks significant performance gains across several domains:
- Dialogue and Recommender Systems: In conversational systems, user intent is fluid, but items exist in a rigid catalog graph. Chen et al. (2023) maximize mutual information to align user-mentioned attributes with graph items, resulting in fewer redundant questions. For open-domain dialogue, Liu et al. (2023) use token-level grounding against KGs to drastically reduce factual hallucinations.
- Fact Verification: Fact-checking requires mapping a nuanced linguistic claim to a hard fact. Bazaga et al. (2024) address this by distilling the semantic knowledge of large language models into structured claim-fact alignments, eliminating the need for expensive human-annotated datasets.
- Graph-Aware Recommendation: Merging sparse user interaction data with dense external knowledge graphs often causes misalignment. Ma (2026) resolves this by using cross-model contrastive learning to reconcile collaborative filtering patterns with multi-relational graph attributes.
By moving away from isolated representations and toward unified, cross-modal architectures, this research direction ensures that models no longer have to choose between the rich context of human language and the unshakeable truth of structured facts.
Key Foundational Work
Neural Knowledge Acquisition via Mutual Attention Between Knowledge Graph and Text (2018) — contributed 1 concept elements to this direction.
Mapping Text to Knowledge Graph Entities using Multi-Sense LSTMs (2018) — contributed 1 concept elements to this direction.
Utilisation de Représentations Distribuées de Relations pour la Désambiguïsation d'Entités Nommées (Exploiting Relation Embeddings to Improve Entity Linking ) (2018) — contributed 1 concept elements to this direction.