Complex Semantic Expression Methods in Voice Assistants: NLP Layers, DIS Limitations, and the CMRL Schema

Introduction – The talk focuses on handling complex semantic expressions in voice assistants, emphasizing that while NLP is mature, understanding intricate text remains challenging.

NLP's Three Levels

Voice assistants process language through three hierarchical stages: lexical analysis, syntactic analysis, and semantic analysis. Lexical analysis converts character strings into tokens (including word segmentation, POS tagging, NER). Syntactic analysis determines grammatical structure or dependency relations, providing essential features for downstream semantic analysis. Semantic analysis builds on the previous layers to infer relationships among linguistic components.

Shallow Semantic Analysis

Shallow analysis, exemplified by Semantic Role Labeling (SRL), identifies predicates and their arguments, assigning semantic roles without constructing a full logical representation.

Semantic Data Structure D.I.S

The widely used D.I.S (Domain‑Intent‑Entity) structure treats a command as a domain (e.g., smart home), an intent (e.g., open), and entities (e.g., place=living room, category=air‑conditioner). While effective for simple commands, D.I.S suffers from six major limitations: domain ambiguity, inability to handle cross‑domain commands, inability to express multi‑entity relations, inability to link multiple intents, inability to capture implicit semantics, and inability to represent ambiguous references.

CMRL: A New Semantic Schema

To overcome D.I.S shortcomings, the CMRL (Complex Meaning Representation Language) schema introduces six element types:

Intent – abstracted predicate nodes.

Thing – object classes.

Enum – enumerated keyword sets.

Operator – relational operators (e.g., >, <, ∈).

Property – attribute edges linking nodes.

Joiner – connectors that express richer relationships beyond simple "is".

These elements can be combined hierarchically to represent complex, multi‑intent, multi‑entity, and ambiguous semantics.

CMRL Advantages

CMRL enables intent reuse across domains, supports cross‑domain commands, captures multi‑entity relations, expresses sequential or parallel intent ordering, represents implicit functions (e.g., brightness), and handles ambiguous references by using abstract concepts such as creativeWork .

Semantic Parsing Algorithms

Transforming natural language into CMRL expressions is treated as a translation problem. Two neural approaches are presented:

Copy‑Write Seq2Seq – The decoder combines a copy mechanism (attending to input tokens) with a write mechanism (selecting schema keywords), drastically reducing the output search space and ensuring relevant tokens are reproduced.

Seq2Tree – A hierarchical tree decoder that generates the logical expression layer‑by‑layer, guaranteeing syntactic correctness by enforcing tree structure and using special span symbols to indicate sub‑nodes.

By integrating these models, the system achieves high‑accuracy parsing of complex voice‑assistant commands.

In summary, the presentation outlines the challenges of complex semantic understanding in voice assistants, critiques the D.I.S model, proposes the CMRL schema, and describes effective neural parsing techniques.