NLG from DMN

Data types

The FEEL data types (p. 106) are:

number, string, boolean, days and time duration, years and months duration, time and date

The time-related data types can be translated into GF functions such as the following:

Duration,   -- {Week,[3..5]}   ~ "Between 3 and 5 weeks"
TimeClock,  -- {Time,14:00}    ~ "At 2 PM", "At any time"
TimeDate    -- {Date,25th May} ~ "On 25th May", "At any date"
 : NP -> FEELexp -> FCell ;

The rest of the types don’t have an obvious semantics in itself. For instance, a numerical type can be a count of some objects, a measurement such as width or height, a serial number, or tons of other things. In order to express the column in natural language, we need to know what the column header is.

WordNet

In the ideal world, I’d try the following:

  • In the XML format, include WordNet identifier.
  • The WordNet identifiers are mapped to GF and to SUMO
  • Use SUMO-to-WordNet mappings to tell us that 05136466 length is mapped to LengthMeasure, and SUMO will tell us how LengthMeasure relates to the rest of the world. Use GF to translate length with the right sense into 10+ languages, with correct inflection table, grammatical gender and such.

However, GF is mapped to WordNet 3.1 and SUMO to WordNet 3.0.

  • TODO: is there a mapping between the versions?

If not, let’s try with only WordNet. There is an ontology there too, just much smaller (fewer relations) than in e.g. SUMO or Cyc.

There is a Haskell library for WordNet (version-agnostic, depends on the actual database files). It has these relations:

-- | The different types of relations which can hold between WordNet Synsets.
data Form = Antonym | Hypernym | Hyponym | Entailment | Similar
          | IsMember | IsStuff | IsPart
          | HasMember | HasStuff | HasPart
          | Meronym | Holonym | CauseTo | PPL | SeeAlso
          | Attribute | VerbGroup | Derivation | Classification | Class | Nominalization
logo