Traditional Spoken Language Understanding (SLU) and semantic parsing systems treat user queries as flat sequences. For a simple command like "Check the weather in Paris," a model can easily map tokens to a single intent (GetWeather) and a set of independent slots (Location: Paris).
However, human language is inherently compositional. Consider a request like: "Get directions to the concert that starts at 8 PM." Here, the destination slot itself contains an entire sub-request requiring its own intent classification and slot filling. As highlighted by Deng et al. in their survey of the field, standard flat architectures suffer a severe performance drop when confronted with this kind of structural complexity.
This limitation is equally pronounced when translating natural language into database queries (Text-to-SQL). While sequence-to-sequence models excel at basic table mappings, Iyer et al. and Guo et al. note that standard models frequently fail to generate deeply nested queries, set operations, or complex relational logic. Attempting to generate a multi-layered logical form in a single, flat pass overtaxes the model's structural reasoning capabilities.
Modeling Meaning as a Tree
To overcome the constraints of flat tagging, researchers shifted toward representing user intent as a hierarchical tree structure. In this paradigm, intents can contain slots, and slots can recursively embed nested intents.
A foundational framework for this approach is the Task Oriented Parsing (TOP) representation introduced by Gupta et al. in Semantic Parsing for Task Oriented Dialog using Hierarchical Representations. Instead of treating an utterance as a linear string of labels, TOP structures it as a grammar tree. For instance, an intent like GetDirections might contain a slot Destination, which recursively contains a nested intent FindEvent and its corresponding slots. This tree-structured representation expands semantic coverage for complex queries by allowing digital assistants to resolve sub-tasks before executing the primary request.
Crucially, this hierarchical framing unifies disparate conversational AI tasks. Goel demonstrates in Practical Semantic Parsing for Spoken Language Understanding that by converting flat intent-slot tags into hierarchical tree structures, both shallow task-oriented commands and deep database queries can be processed within a single, unified semantic parsing framework.
Mechanisms for Hierarchical Decoding
Representing meaning as a tree is only half the battle; the parser must also be capable of generating these structures reliably. Several distinct mechanisms have emerged to handle the combinatorial explosion of nested query generation.
Progressive Decomposition
One intuitive strategy is to break down a complex, deeply nested query into a sequence of simpler, non-nested sub-problems. Liu et al. proposed RECPARSER: A Recursive Semantic Parsing Framework for Text-to-SQL Task, which uses a question decomposer to shift focus across different parts of an utterance, parsing nested clauses incrementally.
Similarly, Guan et al. developed LIST: Linearly Incremental SQL Translator for Single-Hop Reasoning, Generation and Verification. This framework transforms complex text-to-SQL generation into a sequence of verifiable, single-hop reasoning steps, bypassing the need to generate a massive, multi-layered query all at once.
Graph-Based Routing
Rather than relying purely on sequential generation, some approaches explicitly model the structural relationships between parent intents and child slots using graph neural networks. For multi-intent spoken language understanding, Xu et al. introduced a dual relational graph attention network in A BiRGAT Model for Multi-intent Spoken Language Understanding with Hierarchical Semantic Frames. This architecture allows the model to explicitly cluster related semantic slots under their respective parent nodes, resolving the ambiguity that occurs when multiple actions and entities are mixed together in a single spoken utterance.
Reformulation as Question Answering
With the advent of large language models, generating linearized code or complex bracketed syntax directly often leads to structural integrity failures. To mitigate this, Zhao et al. reframed compositional parsing as an abstractive question-answering task in Compositional Task-Oriented Parsing as Abstractive Question Answering. Instead of forcing the model to output a complex tree syntax directly, the system queries the model about specific intents and slots sequentially.
Mekala et al. extended this philosophy to zero-shot settings with ZEROTOP: Zero-Shot Task-Oriented Semantic Parsing using Large Language Models, demonstrating that decomposing a hierarchical parse into a series of unconstrained QA prompts significantly improves structural correctness without requiring domain-specific training data.
Key Foundational Work
Speech intention understanding based on decision tree learning (2004) — contributed 2 concept elements to this direction.
Optimizing user experience through design of the spoken language understanding (SLU) module (2005) — contributed 2 concept elements to this direction.
iSTART: Paraphrase Recognition (2004) — contributed 1 concept elements to this direction.