Zhanliang Liu
EN中文
Back to projects
•
3 min read

FMR: functional meaning representation

Element types in FMR

  • Constants: Refer to specific objects in the world. A constant can be a number, a lexical string, or an entity.
  • Classes: Semantic category of entities. For example: location.city, math.number
    • Sub-class: math.number ⊆ math.expression
    • Template classes: classes with one or more parameters, for example: t.list<c, m, n>, where c is a class, m and n are integers.
  • Functions: The major way to form larger language units from smaller ones. A function is comprised of a name, a list of core arguments, and a return type.
    • Noun functions
      • Map entities to their properties or to other entities having specific relations with the argument(s).
      • Are used to represent noun phrases in natural language.
      • Pronoun functions are special zero-argument noun functions.
    • Verb functions
      • Act as sentences or sub-sentences
      • The most important function type?
    • Modifier functions
      • Many functions can take additional extended arguments as their modifiers.
      • Modifier functions often take the role of extended arguments to modify noun function, verb functions or other modifier functions.
      • Are used in FMR as the semantic representation of adjectives, adverb phrases (including conjunctive adverb phrases), and prepositional phrases in NL
  • Entity Variables: Variables are assigned to FMR nodes for indicating the co-reference of sub-trees to entities.

Features of FMR

  • Strongly typed language
    • Type-compatibility: The type of each child of a function node should match the corresponding argument type of the function.
  • Open-domain type system.
    • Thousands of types
    • Other languages: At most 100+ in grammar level
  • Built-in data structures
    • like t.list and nf.list

Semantic parsing from NL to FMR

By using CFG rules to map natural language sentences and phrases to FMR

"The product of 3 consecutive number is 60" => 
vf.be.equ(nf.math.product(nf.list(math.number, 3, mf.consecutive)), 60);

Origins

FMR grew out of the semantic understanding platform and voice assistant I worked on at Tencent/Sogou from 2013 onward:

  • Compound entity recognition — via CRF, rules, and gazetteers
  • Query intent understanding driven by a language model
  • Template-based query intent recognition and semantic slot extraction
    • A general CFG parser built on an efficient implementation of the Earley algorithm
    • Intent recognition and slot filling
  • Intent ranking, with a final arbitration module over the candidate intents
  • Downstream category-specific knowledge bases and vertical search systems