In a typical conversation, the "topic" isn't a static label; it is a moving target. As we speak, we refer back to previous entities, shift focus between related concepts, and build a mental map of the shared context. Traditional NLP models often struggle with this because they treat dialogue history as a flat, linear sequence of text.
A growing body of research suggests that the key to mastering multi-turn dialogue lies in moving beyond sequences and toward dynamic graphs. By representing a conversation as an evolving network of entities and relationships, models can track the "flow" of a discussion with structural precision that raw text cannot match.
The Bottleneck of Linear History
The standard approach to multi-turn dialogue is to concatenate previous turns into a long string. However, as Yunshi Lan and Jing Jiang argue in their work on Conversational Knowledge Base Question Answering, this method is inherently noisy. It forces the model to sift through irrelevant dialogue tokens to find the "focal entities"—the specific topics that anchor the conversation.
When we say "it" or "that person," we aren't just pointing to a word in a previous sentence; we are pointing to a node in our mental knowledge graph. By explicitly modeling these focal entities and their transitions in a dynamic graph, researchers have found they can resolve coreferences and implicit mentions much more effectively than sequence-based models.
Modeling the Conversation Flow
The core intuition behind this research direction is that a conversation is a process of navigating and updating a graph. Yu Chen and Lingfei Wu pioneered this with GraphFlow, a framework that treats a text passage as a dynamic graph that evolves with each turn. Instead of re-processing the entire text, the model uses a Recurrent Graph Neural Network (RGNN) to pass reasoning states from one question to the next, effectively "flowing" the context forward.
This structural approach is particularly powerful for complex tasks like Conversational Semantic Parsing. Mirella Lapata and Parag Jain demonstrate that by maintaining a dynamic subgraph of a large Knowledge Graph, a system can preserve structural relationships that are lost when dialogue is linearized. This allows the model to "prune" the search space, focusing only on the relevant neighborhood of the graph as the conversation progresses.
Tracking State through Schema Interaction
In task-oriented dialogues—such as booking a hotel or querying a database—the graph isn't just about entities; it's about the schema. Xiaojun Wan and Yitao Cai introduced IGSQL, which uses a database schema interaction graph to track how different tables and columns are referenced across turns.
The challenge with these graphs is that they can be too rigid. Static graphs (like a fixed database schema) don't always capture how a user's goals shift in real-time. To solve this, Yue Feng and colleagues developed DSGFNet, which fuses static domain rules with dynamic, conversation-aware edges. For example, if a user updates a "check-in date" and a "hotel name" simultaneously, the model adds a temporary edge between those slots, capturing a context-specific relationship that didn't exist in the original schema.
Beyond Semantics: Safety and Emotion
The utility of dynamic graphs extends beyond just understanding what a user wants; it also helps in understanding their intent and state.
- Safety: Zixuan Huang and Kecheng Huang developed G-Guard to defend against multi-turn "jailbreak" attacks. By modeling the conversation as an evolving entity-relationship graph, the system can detect when an adversary is incrementally steering an LLM toward harmful behavior—a pattern that is often invisible when looking at turns in isolation.
- Emotion: In Conversational Emotion Recognition, Jiandong Shi and Ming Li use hypergraphs to capture the complex "emotional pathways" between speakers. Because emotions are non-linear and influenced by multiple previous turns, a dynamic graph provides the necessary flexibility to calibrate these relationships as the dialogue heats up or cools down.
The Future of Adaptive Structures
While these models have shown significant gains, the field is moving toward even more autonomy. Early models relied on fixed ontologies, but as Weizhe Lin and Bo-Hsiang Tseng point out in their work on Knowledge-Aware GPT-2, the next frontier is unsupervised structure discovery.
Models like DVAE-GNN, developed by Haifeng Wang and Zeyang Lei, attempt to automatically discover the underlying graph of a conversation from raw data. By learning to map utterances to a hierarchy of topics and semantics without human-defined rules, these systems can adapt to the unique, messy flow of human speech in real-time.