Dynamic Graph Evolution for Conversational Modeling

23 concepts20 papers1 proposal

In conversational AI, tracking context across multiple turns is a fundamental challenge. Standard language models typically linearize dialogue history into a flat sequence of tokens. However, this approach quickly hits token limits and struggles to preserve complex structural relationships, such as long-range coreferences or sudden topic shifts.

An alternative is to use structured representations like database schemas or knowledge graphs. Yet, traditional implementations rely on fixed ontologies or static graphs. As Weizhe Lin et al. point out in their work on Knowledge-Aware Graph-Enhanced GPT-2 for Dialogue State Tracking, a fixed ontology prevents a model from adapting to the unique, unpredictable flow of individual live conversations. Similarly, Haifeng Wang et al. argue in Discovering Dialog Structure Graph for Coherent Dialog Generation that static graph architectures cannot account for emerging conversational trends or unique user inputs encountered during deployment.

To bridge this gap, a powerful paradigm has emerged: dynamic graph evolution. Instead of forcing a conversation into a rigid, pre-defined structure or a flat text stream, these systems model dialogue context as an evolving graph topology that updates in real-time, turn by turn.

Evolving Graph Topologies Across Dialogue Turns

The core intuition behind dynamic graph evolution is to treat dialogue turns, topics, or schema entities as nodes, and dynamically instantiate or modify the edges between them as the interaction unfolds.

For instance, in GraphFlow: Exploiting Conversation Flow with Graph Neural Networks for Conversational Machine Comprehension, Yu Chen et al. construct a dynamic graph over a text passage where nodes and edges evolve at each turn. By using a Recurrent Graph Neural Network (RGNN), the model passes reasoning states from one question to the next, maintaining a context-aware understanding of shifting topics. Similarly, Yi Huang et al. introduce a framework in Modeling Multi-Turn Spoken Language Understanding with Dynamic Graph Convolutional Networks that represents each utterance or discourse turn as a node in an evolving graph. This allows for the adaptive refinement of representations, capturing temporal dependencies far more effectively than standard linear aggregations.

This evolution can also happen over external knowledge bases. In Conversational Semantic Parsing using Dynamic Context Graphs, Mirella Lapata et al. maintain a dynamic subgraph that merges local knowledge graph neighborhoods from consecutive conversation turns. This preserves structural relationships and bypasses the token limits associated with long text sequences. As highlighted by Kang Liu et al. in the Editorial for Special Issue on Commonsense Knowledge and Reasoning: Representation, Acquisition and Applications, combining recurrent graph networks with dynamic graph structures allows systems to maintain vital temporal dependencies across multiple interactions.

Real-Time Schema and Relationship Augmentation

Beyond tracking dialogue history, dynamic graph evolution is highly effective for context-dependent tasks like Dialogue State Tracking (DST) and Text-to-SQL generation, where user intent must be mapped to structured databases.

In multi-domain DST, a user might transition from booking a hotel to reserving a restaurant, reusing information like the number of guests or the date. Yue Feng et al. address this in Dynamic Schema Graph Fusion Network for Multi-Domain Dialogue State Tracking by dynamically adding edges to a static schema graph during a conversation. These new edges represent real-time relationships, such as co-reference and co-update events (e.g., when multiple slots are updated simultaneously). This conversation-specific schema augmentation allows the system to adapt its internal representation of slot relationships on the fly, a strategy that Su Zhu et al. demonstrate in Efficient Context and Schema Fusion Networks for Multi-Domain Dialogue State Tracking to be highly effective at combating data sparsity.

A similar mechanism is used in semantic parsing. Xiaojun Wan et al. introduce IGSQL: Database Schema Interaction Graph Based Neural Model for Context-Dependent Text-to-SQL Generation, which models database schemas as dynamic graphs with cross-turn layers. This propagates information between related entities over time, tracking active neighborhoods within the schema graph. To prevent older, irrelevant context from polluting the current turn, Binyuan Hui et al. employ a "memory decay" mechanism in Dynamic Hybrid Relation Exploration Network for Cross-Domain Context-Dependent Semantic Parsing, updating the graph dynamically to capture the evolving relationship between user intent and relational data structures while letting obsolete history fade.

Even in early robotics work, Stefanie Tellex et al. demonstrated in Toward Information Theoretic Human-Robot Dialog that merging probabilistic graphs of initial commands and subsequent responses via coreference resolution allows a system to update its internal world model incrementally without restarting inference.

Why It Works: Intuition and Information Routing

The fundamental reason dynamic graph evolution outperforms static or linear baselines comes down to selective information routing. In a static graph or a flat sequence, attention or message-passing steps are distributed uniformly or constrained by fixed topologies. In contrast, dynamic graphs adjust edge weights or create temporary connections based on real-time feedback and semantic similarity, focusing the model's computational capacity precisely where it matters.

Consider conversational recommendation systems. In Unified Conversational Recommendation Policy Learning via Graph-based Reinforcement Learning, Yang Deng et al. represent the conversational state using a graph of users, items, and attributes with dynamic edge weights. By updating these weights in real-time based on direct user feedback, the system effectively re-ranks the entire item space at every turn. Similarly, Xuan Ma et al. develop the Dynamic Open-book Prompt for Conversational Recommender System, which dynamically modifies its internal graph structure by instantiating temporary edges between the current user utterance and relevant historical items based on semantic similarity. An utterance-aware graph convolutional layer then aggregates these dynamic connections to filter and prioritize information.

This real-time adaptation closely mimics human cognitive patterns. As Junjie Wu et al. demonstrate in Augmenting Topic Aware Knowledge-Grounded Conversations with Dynamic Built Knowledge Graphs, constructing a dynamic knowledge graph from dialogue history allows a system to model logical relationships between topics as they emerge on the fly, leading to more coherent transitions. Finally, in complex domains like emotion recognition, Jiandong Shi et al. note in MATCH: Modality-Calibrated Hypergraph Fusion Network for Conversational Emotion Recognition that moving away from fixed connectivity patterns toward adaptive structures allows models to dynamically adjust the strength of relationships as emotional pathways and dialogue contexts shift.

By treating context not as a static backdrop or a simple text stream, but as a living, breathing network, dynamic graph evolution enables conversational agents to maintain structural clarity, adapt to user feedback, and scale gracefully across complex, multi-turn interactions.

Key Foundational Work

GraphFlow: Exploiting Conversation Flow with Graph Neural Networks for Conversational Machine Comprehension (2020) — contributed 1 concept elements to this direction.
Efficient Context and Schema Fusion Networks for Multi-Domain Dialogue State Tracking (2020) — contributed 1 concept elements to this direction.
IGSQL: Database Schema Interaction Graph Based Neural Model for Context-Dependent Text-to-SQL Generation (2020) — contributed 1 concept elements to this direction.
History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 19 papers
  2. Version 1 · 2026-10-04 · 16 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_dynamic_schema_modeling_8950de57,
  title = {dynamic schema modeling},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/dynamic-schema-modeling/rdr_8950de57071d4b75a9347aaf1ef3392d},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-22},
}

This exact version (version 2, frozen text):

@misc{lacuna_dynamic_schema_modeling_8950de57_v2,
  title = {dynamic schema modeling},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/dynamic-schema-modeling/rdr_8950de57071d4b75a9347aaf1ef3392d/v/2},
  note = {Version 2, frozen 2026-10-04. Accessed 2026-11-22},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake