In natural conversation, we rarely speak in isolated fragments. Instead, our utterances are nested within larger strategic units. A simple "Okay" might be a linguistic move to signal understanding, which sits inside an exchange to confirm a fact, which itself is part of a broader transaction to book a flight. Research into hierarchical dialogue modeling treats conversation not as a flat sequence of turns, but as a structured tree where high-level goals are decomposed into increasingly granular sub-tasks.
The Isomorphism of Task and Talk
The central intuition behind hierarchical modeling is that task-oriented conversations inherit their organization from the task itself. As Barbara J. Grosz et al. argued in Using Process Knowledge in Understanding Task-Oriented Dialogs, the progression of a collaborative project dictates the boundaries of sub-dialogues. If a task has a natural hierarchy—such as assembling a machine or solving a technical problem—the conversation will mirror that structure.
Raymonde Guindon and Joyce Conner refer to this as an isomorphism between dialogue and task goals in The Structure of User-Adviser Dialogues: Is there Method in their Madness?. By mapping surface-level utterances back to a hierarchy of motivations and strategies, as proposed by Martha E. Pollack et al. in User Participation in the Reasoning Processes of Expert Systems, systems can track a user's high-level objectives even when the immediate linguistic move seems ambiguous.
Formalizing the Conversational Tree
To represent this hierarchy computationally, researchers have developed formal grammars for discourse. Tom Wachtel introduced a tree-based representation in Pragmatic sensitivity in NL interfaces and the structure of conversation, using rewrite rules to define how a conversation (CONV) decomposes into dialogues (DIAL) and exchanges (EXCH).
This decomposition typically follows a standard taxonomy:
- Transactions: The highest level, representing a complete task or narrative stage.
- Exchanges: Mid-level units where participants interact to achieve a specific sub-goal (e.g., a question-answer pair).
- Moves/Acts: The atomic level, representing the functional intent of a single utterance.
Eric Bilange and Arne Jönsson both utilized these nested units to manage conversational flow, allowing systems to handle complex phenomena like initiative shifts and sub-dialogues without losing track of the primary goal.
Signaling and Tracking Structure
Humans use various cues to signal where they are in this hierarchy. James Raymond Davis and Julia Hirschberg demonstrated in Assigning Intonational Features in Synthesized Spoken Directions that intonation changes based on whether a speaker is starting a new high-level segment or continuing a sub-task. Similarly, Justine Cassell et al. showed that physical posture shifts serve as "visual punctuation" for topic transitions in Non-Verbal Cues for Discourse Structure.
To track these transitions, systems often employ a focus stack. As participants move into a sub-task, the new discourse segment is pushed onto the stack; once the sub-task is completed, it is popped, returning the "focus" to the parent goal. This mechanism is essential for resolving references and understanding which part of a complex instruction is currently relevant.
Benefits for Machine Learning
Modern neural approaches have validated these classical structural insights. Kai Cong and Wai Lam found that Hierarchical Attention Networks, which explicitly model the relationship between sentences and turns, outperform flat models in evaluating dialogue quality. By reflecting the natural hierarchy of speech—from words to utterances to full dialogues—models like those proposed by Emile Chapuis et al. can capture discourse-level features that flat sequence models overlook.
Furthermore, accounting for this structure improves performance prediction. Mihai Rotaru and Diane J. Litman showed in Exploiting Discourse Structure for Spoken Dialogue Performance Analysis that models considering how a dialogue moves between main tasks and sub-tasks explain significantly more variance in system success than those using flat interaction metrics. While Srinivas Bangalore and Amanda Stent note that simple linear models are often sufficient for basic tasks, the hierarchical "parse-based" view remains necessary for capturing the true complexity of human-like, nested communication.
Key Foundational Work
Modelling Simple Dialogs (1977) — contributed 1 concept elements to this direction.
User Participation in the Reasoning Processes of Expert Systems (1982) — contributed 1 concept elements to this direction.
Using Process Knowledge in Understanding Task-Oriented Dialogs (1977) — contributed 1 concept elements to this direction.