Naturalistic DSLs

This document is based on Chapter 7. Discussion from the PhD Thesis:

“Design and Implementation of Effect Handlers for Object-Oriented Languages”
Jonathan Immanuel Brachthäuser, Universität Tübingen (2019).

(We translated it to the Effekt language, revised it and added more explanations and additional examples.)

It is the goal of domain specific languages (DSLs) to bridge the conceptual gap between language used in the particular domain and the computer language domain experts use to solve the domain problems (Hudak, 1996). Domain language is often close to natural (that is, spoken) language, making the task of DSL design particularly challenging (Lopes et al., 2003). Natural language has the reputation of being lexically and syntactically ambiguous, having complicated and context dependent binding structures, and often has non-trivial semantics, which rarely is compositional.

In consequence, often DSLs are still far away from being close to natural language. This is in particular the case for DSLs, which are embedded (Hudak, 1996) into a general-purpose language. With embedded DSLs, the host language additionally imposes its own syntactical restrictions and typing discipline on the DSL designer. Many limitations have been addressed in their own line of work. Examples include syntax extensions as libraries (Erdweg et al., 2011) and domain specific type system extensions. However, non-context-free linguistic constructs are often neglected.

In about the last decade, many developments in modeling the semantics of natural languages have been inspired by computer science and the theory of abstract machines and control operators in particular. Importantly, Chung-chieh Shan describes “noncompositional phenomena in natural languages” as linguistic side effects. Delimited continuations have successfully been used to model linguistic side effects, such as quantification, focus, and polymorphic coordination (Shan, 2004 and 2005; Barker, 2004). Maršík and Amblard recently used algebraic effects with handlers to give a compositional semantics to deixis (“John loves *me*“), quantification with scope islands (“John loves *every woman*“), and implicature (“John, *my best friend*, loves me”).

Implementors of (domain specific) programming languages often reside to effects to describe the semantics of the language. Inspired by the recent developments in natural language semantics, here we propose to follow Maršík and Amblard and use effects and handlers to describe the syntax of a DSL. In particular, we group the different syntactic constructs of a DSL according to the following aspects:

  1. pure syntax, that can be understood as compositional construction of the abstract syntax tree.
  2. effectful syntax, that, like linguistic side effects (Shan, 2005), (often) requires context for interpretation and results in some form of non-local rewriting of the syntax tree. Effect operations can be used to express effectful syntax.
  3. binding syntax, which provides the necessary context. Binding syntax can be expressed as effect handlers for effectful syntax.

Effectful Syntax in Effekt

Effekt comes equipped with a few features that allow designing elegant embedded DSLs. One example is the infix notation of function application. Another one, as we will see, is effect handlers, which we can use to implement effectful DSLs. To illustrate the gained expressivity, we implement examples from Maršík and Amblard as an embedded domain specific language. Maršík and Amblard already use a calculus of effects and handlers to express the semantics. We simply translate the examples to Effekt.

The Language of Sentence

We start by describing the language (DSL) of nominal phrases as the following datatype:

module examples/casestudies/naturalisticdsls

record Person(name: String)

Let us define some example people:

val John = Person("John")
val Peter = Person("Peter")
val Mary = Person("Mary")

Next, we define the sentence DSL:

type Sentence {
  Say(speaker: Person, sentence: Sentence)
  Is(person: Person, predicate: Predicate)
   // used later:
  ForAll(individual: Person, s: Sentence)
  Implies(a: Sentence, b: Sentence)
}
type Predicate {
  InLoveWith(p: Person)
  Woman()
}
def loves(lover: Person, loved: Person) = Is(lover, InLoveWith(loved))

We can now construct sentences like

John said “Mary loves me”

as:

val s1 = Say(John, Mary.loves(John))

The Speaker Effect

However, there is a better way to express this: Using effects. We define the speaker effect to refer to the contextual speaker of the sentence.

interface Speaker {
  def speaker(): Person
}
def me() = do speaker()

The semantics of me depends on the context. If we treat quotes in the above sentence as scope, we see that “said” handles me. That is:

def said(p: Person) { s: => Sentence / Speaker }: Sentence / {} =
  try { Say(p, s()) } with Speaker { def speaker() = resume(p) }

Here is another definition of said, that does not handle the speaker effect, corresponding to omitting the quotationmarks:

def said(p: Person, s: Sentence): Sentence =
  Say(p, s)

We can now express our example sentence as:

def s1a() = John.said { Mary.loves(me()) }

Note, that by overloading said we can also simply express the sentence

John said Mary loves me

as

def s1b() = John.said ( Mary.loves(me()) )

Comparing the inferred types of s1a and s1b we see:

def s1aTpe(): Sentence / {} = s1a()
def s1bTpe(): Sentence / Speaker = s1b()

That is, s1b still has the speaker effect, while s1a is pure and all linguistic effects are handled. We can handle s1b by adding the speaker:

def s1c(): Sentence / {} = Peter.said { s1b() }

which results in

//> Say(Person(Peter),
//    Say(Person(John),
//      Is(Person(Mary), InLoveWith(Person(Peter)))))

This simple example already illustrates that:

  • effects can be used to model contextual information in natural language
  • the effect system reminds us to actually handle (that is, provide semantics to) linguistic effects

The Scope effect

Passing down context information, as we did with the speaker effect, does not require full handlers. In Scala, for example, implicit parameters would suffice to express this effect. Things become more interesting when we consider the scope effect, which can be used to model universal quantification (Maršík and Amblard, 2016).

Instead of modeling scope directly as an effect, here we treat quantification as the effect:

interface Quantification {
  def quantify(who: Predicate): Person
}
def every(who: Predicate) = do quantify(who)

We can use it as follows:

def s2(): Sentence = scoped { John.said { every(Woman()).loves(me()) } }

The effect operation every takes a predicate (&ie;, Woman) and introduces a universal quantification at the position of the handler scoped.

The quantification effect is handled by scoped:

def scoped { s: => Sentence / Quantification }: Sentence = {
  var tmp = 0;
  def fresh(): Person = { val x = Person("x" ++ show(tmp)); tmp = tmp + 1; x }
  try { s() } with Quantification {
    def quantify(who) = {
      val x = fresh()
      ForAll(x, Implies(Is(x, who), resume(x)))
    }
  }
}

This is already a more involved handler. It generates a fresh person name and then systematically rewrites the syntax tree, moving the introduced binder and the predicate up to the handler:

//> ForAll(Person(x0), Implies(Is(Person(x0), Woman()),
//    Say(Person(John), Is(Person(x0), InLoveWith(Person(John))))))

Every invocation of the effect operation every introduces an additional quantifier. This non-local rewriting of the syntax tree to introduce a binder is very similar to let-insertion (Yallop, 2017). Yallop shows how to use effect handlers to perform let-insertion.

Running the Examples

Finally, we can run our examples to inspect the generated sentences.

def main() = {
  println(s1)
  println(s1a())
  println(s1c())
  println(s2())
}