Graph Based Dialogue Representation

743 concepts397 papers6 proposals

When humans engage in conversation, we rarely think in a strictly linear fashion. A comment made now might directly answer a question from ten turns ago, completely bypassing the intervening small talk. In multi-party chats, several sub-conversations often run in parallel, thoroughly entangled in time.

Traditional language models treat dialogue as a flat, chronological sequence of text tokens. While this sequential paradigm works well for structured prose, it struggles with the non-linear, multi-threaded nature of human conversation. Forcing a model to learn these complex relationships via standard sequential attention introduces significant noise and requires the model to bridge massive temporal gaps.

To address this, a prominent research direction models conversations as relational graphs. By transforming a dialogue into a network where utterances, speakers, and entities serve as nodes, and conversational dependencies serve as edges, models can directly capture the underlying topology of human interaction.

Constructing the Dialogue Graph

The core intuition behind this approach is that the structural layout of a conversation is just as informative as the text itself. To capture this structure, researchers construct graphs that explicitly map out how different components of a dialogue interact.

[Utterance 1: Speaker A] --(Intra-Speaker Edge)--> [Utterance 3: Speaker A]
          \                                               ^
           \                                             /
     (Inter-Speaker Reply)                         (Coreference Link)
             \                                         /
              v                                       /
        [Utterance 2: Speaker B] --------------------/

Depending on the task, these graphs incorporate different types of nodes and edges:

  • Speaker and Utterance Dynamics: To track how individuals influence one another, graphs can separate speaker identity from the text. Ghosal et al. (2019) in their DialogueGCN framework represent utterances as nodes and categorize edges by speaker identity and temporal direction. This allows the model to distinguish between a speaker's internal emotional transitions (intra-speaker inertia) and the external influence of other participants (inter-speaker influence). Similarly, Chen Zhang et al. (2021) utilize speaker-allocated relational edges to evaluate global dialogue dynamics and detect structural inconsistencies.
  • Discourse and Functional Relations: Conversations have functional flow. Dong Wang et al. (2011) demonstrate that linking utterances via dialogue acts—such as question-answer pairs or agreement links—allows importance scores to propagate naturally through a conversation. In educational dialogues, Jain et al. (2023) use Levi Graphs to map pedagogical intentions, tracking how specific tutor hints lead directly to student affirmations.
  • Linguistic and Semantic Cross-Links: Beyond turn-level connections, graphs can map fine-grained semantic dependencies. Zhengyuan Liu et al. (2021) connect entity mentions across distant turns using coreference edges, creating an explicit pipeline for information flow around a single subject. Taking this further, the authors of a study on semantic representation unify sentence-level Abstract Meaning Representation (AMR) graphs into a global dialogue-level graph, linking identical concepts and speakers across the entire history.

Why Graph Representations Work

Shifting from a linear sequence to a graph structure fundamentally alters how neural networks process conversational context. It provides three distinct architectural advantages:

1. Bypassing Sequential Noise

In multi-party chats, side-conversations act as structural noise for the main thread. Kummerfeld et al. (2019) frame conversation disentanglement as a graph-prediction task where messages are nodes and "reply-to" relationships are directed edges, treating independent conversations as disconnected components.

When reading these complex threads, graph neural networks (GNNs) can bypass irrelevant intervening turns entirely. As Jiaqi Li et al. (2021) demonstrate with DADgraph, routing information through functional discourse links rather than chronological order allows a model to directly access pertinent context located dozens of turns away, mitigating the noise of overlapping conversations.

2. Explicit Information Propagation

Graphs allow models to mathematically transfer context or significance between related ideas. For instance, in conversational summarization, an informative answer might lack explicit topic keywords, making it look unimportant to a standard sequential model. By using a graph-based ranking approach, Dong Wang et al. (2011) show that the structural link between a question and an answer allows the topical importance of a query to flow directly into its response.

3. Capturing Long-Range Trajectories

Human emotional and topical states persist over time. Jiaxing Liu et al. (2021) show that using graph convolutions to aggregate information from neighboring conversational nodes smooths out local variations and highlights broader emotional trajectories.

To capture deep psychological context, Jiangnan Li et al. (2021) utilize a Graph Transformer to map a speaker's past actions, current effects, and future intentions as explicit structural paths, allowing the model to recognize psychological influences that resurface across long temporal gaps.

Advanced Topologies and Heterogeneous Networks

As the field has matured, the structure of these dialogue graphs has evolved from simple turn-to-turn chains into complex, multi-layered, and heterogeneous networks.

Heterogeneous Node Types

Instead of treating every node as a text snippet, models increasingly mix modalities and abstractions. Dong Zhang et al. (2019) introduce ConGCN, which models both utterances and speakers as distinct node types within a single heterogeneous graph. Message passing on this graph allows an utterance node to simultaneously integrate local conversational context from neighboring utterance nodes and historical personality traits from its corresponding speaker node.

Grounding with External Knowledge

Dialogue graphs are frequently paired with external knowledge bases to handle complex reasoning tasks. For multi-domain task-oriented dialogues, Libo Qin et al. (2023) establish dynamic edges between conversation tokens and entities in an external knowledge graph based on co-occurrence, allowing the model to navigate multiple databases simultaneously. For knowledge-grounded generation, Kexin Wang et al. (2022) construct a specialized subgraph that maps the conversation's progression as a physical path across an external knowledge graph, ensuring that topic transitions remain logically consistent.

Hierarchical and Stacked Graphs

For highly complex multimodal settings, such as Visual Dialog, models often stack multiple specialized graphs. Feilong Chen et al. (2021) design a Graph-over-Graph (GoG) network that layers three distinct structures: a dialogue history coreference graph, a question dependency graph, and an image spatial relation graph. This hierarchical architecture allows the model to resolve linguistic ambiguities in the text graph before projecting those entities into the visual graph to locate the answer.

By moving away from the constraints of linear text processing, graph-based dialogue modeling provides a mathematically rigorous way to capture the true, non-linear essence of human communication. Whether tracking emotions, generating summaries, or extracting medical data, representing conversations as topological networks ensures that models reason with structure, rather than just sequence.

History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_graph_based_conversation_modeling_710e6c83,
  title = {graph-based conversation modeling},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/graph-based-conversation-modeling/rdr_710e6c83e8254e0aa108d95978cfb07d},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_graph_based_conversation_modeling_710e6c83_v1,
  title = {graph-based conversation modeling},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/graph-based-conversation-modeling/rdr_710e6c83e8254e0aa108d95978cfb07d/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake