[[!toc]] Monads ------ Start by (re)reading the discussion of monads in the lecture notes for week 6 [[Towards Monads]]. In those notes, we saw a way to separate thinking about error conditions (such as trying to divide by zero) from thinking about normal arithmetic computations. We did this by making use of the option monad: in each place where we had something of type `int`, we put instead something of type `int option`, which is a sum type consisting either of just an integer, or else some special value which we could interpret as signaling that something had gone wrong. The goal was to make normal computing as convenient as possible: when we're adding or multiplying, we don't have to worry about generating any new errors, so we do want to think about the difference between `int`s and `int option`s. We tried to accomplish this by defining a `bind` operator, which enabled us to peel away the `option` husk to get at the delicious integer inside. There was also a homework problem which made this even more convenient by mapping any binary operation on plain integers into a lifted operation that understands how to deal with `int option`s in a sensible way. [Linguitics note: Dividing by zero is supposed to feel like a kind of presupposition failure. If we wanted to adapt this approach to building a simple account of presupposition projection, we would have to do several things. First, we would have to make use of the polymorphism of the `option` type. In the arithmetic example, we only made use of `int option`s, but when we're composing natural language expression meanings, we'll need to use types like `N option`, `Det option`, `VP option`, and so on. But that works automatically, because we can use any type for the `'a` in `'a option`. Ultimately, we'd want to have a theory of accommodation, and a theory of the situations in which material within the sentence can satisfy presuppositions for other material that otherwise would trigger a presupposition violation; but, not surprisingly, these refinements will require some more sophisticated techniques than the super-simple option monad.] So what exactly is a monad? As usual, we're not going to be pedantic about it, but for our purposes, we can consider a monad to be a system that provides at least the following three elements: * A complex type that's built around some more basic type. Usually it will be polymorphic, and so can apply to different basic types. In our division example, the polymorphism of the `'a option` type provides a way of building an option out of any other type of object. People often use a container metaphor: if `x` has type `int option`, then `x` is a box that (may) contain an integer. type 'a option = None | Some of 'a;; * A way to turn an ordinary value into a monadic value. In OCaml, we did this for any integer `n` by mapping it to the option `Some n`. In the general case, this operation is known as `unit` or `return.` Both of those names are terrible. This operation is only very loosely connected to the `unit` type we were discussing earlier (whose value is written `()`). It's also only very loosely connected to the "return" keyword in many other programming languages like C. But these are the names that the literature uses. The unit/return operation is a way of lifting an ordinary object into the monadic box you've defined, in the simplest way possible. You can think of the singleton function as an example: it takes an ordinary object and returns a set containing that object. In the example we've been considering: let unit x = Some x;; val unit : 'a -> 'a option = So `unit` is a way to put something inside of a monadic box. It's crucial to the usefulness of monads that there will be monadic boxes that aren't the result of that operation. In the option/maybe monad, for instance, there's also the empty box `None`. In another (whimsical) example, you might have, in addition to boxes merely containing integers, special boxes that contain integers and also sing a song when they're opened. The unit/return operation will always be the simplest, conceptually most straightforward way to lift an ordinary value into a monadic value of the monadic type in question. * Thirdly, an operation that's often called `bind`. This is another unfortunate name: this operation is only very loosely connected to what linguists usually mean by "binding." In our option/maybe monad, the bind operation is: let bind m f = match m with None -> None | Some n -> f n;; val bind : 'a option -> ('a -> 'b option) -> 'b option = Note the type. `bind` takes two arguments: first, a monadic "box" (in this case, an 'a option); and second, a function from ordinary objects to monadic boxes. `bind` then returns a monadic value: in this case, a 'b option (you can start with, e.g., int options and end with bool options). Intuitively, the interpretation of what `bind` does is like this: the first argument is a monadic value m, which evaluates to a box that (maybe) contains some ordinary value, call it `x`. Then the second argument uses `x` to compute a new monadic value. Conceptually, then, we have let bind m f = (let x = unbox m in f x);; The guts of the definition of the `bind` operation amount to specifying how to unbox the monadic value `m`. In the bind opertor for the option monad, we unboxed the option monad by matching the monadic value `m` with `Some n`---whenever `m` happened to be a box containing an integer `n`, this allowed us to get our hands on that `n` and feed it to `f`. If the monadic box didn't contain any ordinary value, then we just pass through the empty box unaltered. In a more complicated case, like our whimsical "singing box" example from before, if the monadic value happened to be a singing box containing an integer `n`, then the `bind` operation would probably be defined so as to make sure that the result of `f n` was also a singing box. If `f` also inserted a song, you'd have to decide whether both songs would be carried through, or only one of them. There is no single `bind` function that dictates how this must go. For each new monadic type, this has to be worked out in an useful way. So the "option/maybe monad" consists of the polymorphic option type, the unit/return function, and the bind function. With the option monad, we can think of the "safe division" operation
# let divide' num den = if den = 0 then None else Some (num/den);;
val divide' : int -> int -> int option = 
as basically a function from two integers to an integer, except with this little bit of option plumbing on the side. A note on notation: Haskell uses the infix operator `>>=` to stand for `bind`. Chris really hates that symbol. Following Wadler, he prefers to use an infix five-pointed star, or on a keyboard, `*`. Jim on the other hand thinks `>>=` is what the literature uses and students won't be able to avoid it. Moreover, although ⋆ is OK (though not a convention that's been picked up), overloading the multiplication symbol invites its own confusion and Jim feels very uneasy about that. If not `>>=` then we should use some other unfamiliar infix symbol (but `>>=` already is such...) In any case, the course leaders will work this out somehow. In the meantime, as you read around, wherever you see `m >>= f`, that means `bind m f`. Also, if you ever see this notation: do x <- m f x That's a Haskell shorthand for `m >>= (\x -> f x)`, that is, `bind m f`. Similarly: do x <- m y <- n f x y is shorthand for `m >>= (\x -> n >>= (\y -> f x y))`, that is, `bind m (fun x -> bind n (fun y -> f x y))`. Those who did last week's homework may recognize this. (Note that the above "do" notation comes from Haskell. We're mentioning it here because you're likely to see it when reading about monads. It won't work in OCaml. In fact, the `<-` symbol already means something different in OCaml, having to do with mutable record fields. We'll be discussing mutation someday soon.) As we proceed, we'll be seeing a variety of other monad systems. For example, another monad is the list monad. Here the monadic type is: # type 'a list The unit/return operation is: # let unit x = [x];; val unit : 'a -> 'a list = That is, the simplest way to lift an 'a into an 'a list is just to make a singleton list of that 'a. Finally, the bind operation is: # let bind m f = List.concat (List.map f m);; val bind : 'a list -> ('a -> 'b list) -> 'b list = What's going on here? Well, consider (List.map f m) first. This goes through all the members of the list m. There may be just a single member, if `m = unit a` for some a. Or on the other hand, there may be no members, or many members. In any case, we go through them in turn and feed them to f. Anything that gets fed to f will be an 'a. f takes those values, and for each one, returns a 'b list. For example, it might return a list of all that value's divisors. Then we'll have a bunch of 'b lists. The surrounding `List.concat ( )` converts that bunch of 'b lists into a single 'b list: # List.concat [[1]; [1;2]; [1;3]; [1;2;4]] - : int list = [1; 1; 2; 1; 3; 1; 2; 4] So now we've seen two monads: the option/maybe monad, and the list monad. For any monadic system, there has to be a specification of the complex monad type, which will be parameterized on some simpler type 'a, and the unit/return operation, and the bind operation. These will be different for different monadic systems. Many monadic systems will also define special-purpose operations that only make sense for that system. Although the unit and bind operation are defined differently for different monadic systems, there are some general rules they always have to follow. The Monad Laws -------------- Just like good robots, monads must obey three laws designed to prevent them from hurting the people that use them or themselves. * **Left identity: unit is a left identity for the bind operation.** That is, for all `f:'a -> 'a m`, where `'a m` is a monadic value, we have `(unit x) * f == f x`. For instance, `unit` is itself a function of type `'a -> 'a m`, so we can use it for `f`:
# let ( * ) m f = match m with None -> None | Some n -> f n;;
val ( * ) : 'a option -> ('a -> 'b option) -> 'b option = 
# let unit x = Some x;;
val unit : 'a -> 'a option = 

# unit 2;;
- : int option = Some 2
# unit 2 * unit;;
- : int option = Some 2

# divide 6 2;;
- : int option = Some 3
# unit 2 * divide 6;;
- : int option = Some 3

# divide 6 0;;
- : int option = None
# unit 0 * divide 6;;
- : int option = None
The parentheses is the magic for telling OCaml that the function to be defined (in this case, the name of the function is `*`, pronounced "bind") is an infix operator, so we write `m * f` or `( * ) m f` instead of `* m f`. * **Associativity: bind obeys a kind of associativity**. Like this: (m * f) * g == m * (fun x -> f x * g) If you don't understand why the lambda form is necessary (the "fun x" part), you need to look again at the type of bind. Some examples of associativity in the option monad:
# Some 3 * unit * unit;; 
- : int option = Some 3
# Some 3 * (fun x -> unit x * unit);;
- : int option = Some 3

# Some 3 * divide 6 * divide 2;;
- : int option = Some 1
# Some 3 * (fun x -> divide 6 x * divide 2);;
- : int option = Some 1

# Some 3 * divide 2 * divide 6;;
- : int option = None
# Some 3 * (fun x -> divide 2 x * divide 6);;
- : int option = None
Of course, associativity must hold for arbitrary functions of type `'a -> 'a m`, where `m` is the monad type. It's easy to convince yourself that the bind operation for the option monad obeys associativity by dividing the inputs into cases: if `m` matches `None`, both computations will result in `None`; if `m` matches `Some n`, and `f n` evalutes to `None`, then both computations will again result in `None`; and if the value of `f n` matches `Some r`, then both computations will evaluate to `g r`. * **Right identity: unit is a right identity for bind.** That is, `m * unit == m` for all monad objects `m`. For instance,
# Some 3 * unit;;
- : int option = Some 3
# None * unit;;
- : 'a option = None
More details about monads ------------------------- If you studied algebra, you'll remember that a *monoid* is an associative operation with a left and right identity. For instance, the natural numbers along with multiplication form a monoid with 1 serving as the left and right identity. That is, temporarily using `*` to mean arithmetic multiplication, `1 * n == n == n * 1` for all `n`, and `(a * b) * c == a * (b * c)` for all `a`, `b`, and `c`. As presented here, a monad is not exactly a monoid, because (unlike the arguments of a monoid operation) the two arguments of the bind are of different types. But it's possible to make the connection between monads and monoids much closer. This is discussed in [[Advanced Notes/Monads in Category Theory]]. See also . Here are some papers that introduced monads into functional programming: * [Eugenio Moggi, Notions of Computation and Monads](http://www.disi.unige.it/person/MoggiE/ftp/ic91.pdf): Information and Computation 93 (1). * [Philip Wadler. Monads for Functional Programming](http://homepages.inf.ed.ac.uk/wadler/papers/marktoberdorf/baastad.pdf): in M. Broy, editor, *Marktoberdorf Summer School on Program Design Calculi*, Springer Verlag, NATO ASI Series F: Computer and systems sciences, Volume 118, August 1992. Also in J. Jeuring and E. Meijer, editors, *Advanced Functional Programming*, Springer Verlag, LNCS 925, 1995. Some errata fixed August 2001. The use of monads to structure functional programs is described. Monads provide a convenient framework for simulating effects found in other languages, such as global state, exception handling, output, or non-determinism. Three case studies are looked at in detail: how monads ease the modification of a simple evaluator; how monads act as the basis of a datatype of arrays subject to in-place update; and how monads can be used to build parsers. * [Philip Wadler. The essence of functional programming](http://homepages.inf.ed.ac.uk/wadler/papers/essence/essence.ps): invited talk, *19'th Symposium on Principles of Programming Languages*, ACM Press, Albuquerque, January 1992. This paper explores the use monads to structure functional programs. No prior knowledge of monads or category theory is required. Monads increase the ease with which programs may be modified. They can mimic the effect of impure features such as exceptions, state, and continuations; and also provide effects not easily achieved with such features. The types of a program reflect which effects occur. The first section is an extended example of the use of monads. A simple interpreter is modified to support various extra features: error messages, state, output, and non-deterministic choice. The second section describes the relation between monads and continuation-passing style. The third section sketches how monads are used in a compiler for Haskell that is written in Haskell. There's a long list of monad tutorials on the [[Offsite Reading]] page. Skimming the titles makes me laugh. In the presentation we gave above, which is the usual one in functional programming, we took unit/return and bind as the primitive operations. From these a number of other general monad operations can be derived. It's also possible to take some of the others as primitive, and define for example bind in terms of them. This is discussed some in the [[Advanced Notes/Monads in Category Theory]]. Here are some of the specifics. You don't have to master these; they're collected here for your reference. You may sometimes see: m >> n That just means: m >>= fun _ -> n that is: bind m (fun _ -> n) You could also do `bind m (fun x -> n)`; we use the `_` for the function argument to be explicit that that argument is never going to be used. The `lift` operation we asked you to define for last week's homework is a common operation. The second argument to `bind` converts 'a values into `b m values---that is, into instances of the monadic type. What if we instead had a function that merely converts `a values into `b values, and we want to use it with our monadic type. Then we "lift" that function into an operation on the monad. For example: # let even x = (x mod 2 = 0);; val g : int -> bool = `even` has the type int -> bool. Now what if we want to convert it into an operation on the option/maybe monad? # let lift g = fun m -> bind m (fun x -> Some (g x));; val lift : ('a -> 'b) -> 'a option -> 'b option = `lift even` will now be a function from `int option`s to `bool option`s. We can also define a lift operation for binary functions: # let lift2 g = fun m n -> bind m (fun x -> bind n (fun y -> Some (g x y)));; val lift2 : ('a -> 'b -> 'c) -> 'a option -> 'b option -> 'c option = `lift (+)` will now be a function from `int option`s and `int option`s to `int option`s. This should look familiar to those who did the homework. The `lift` operation (just `lift`, not `lift2`) is sometimes also called the `map` operation. (In Haskell, they say `fmap` or `<$>`.) And indeed when we're working with list monad, `lift f` is exactly `List.map f`! Wherever we have a well-defined monad, we can define a lift/map operation for that monad. The examples above used `Some (g x)` and so on; in the general case we'd use `unit (g x)`, using the pre-defined `unit` operation for the monad we're working with. In general, any lift/map operation can be relied on to satisfy these laws: * lift id = id * lift (compose f g) = compose (lift f) (lift g) where id is `fun x -> x` and `compose a b` is `fun x -> a (b x)`. If you think about the special case of the map operation on lists, this should make sense. `List.map id lst` should give you back `lst` again. And you'd expect these two computations to give the same result: List.map (fun x -> f (g x)) lst List.map f (List.map g lst) Another general monad operation is called `ap` in Haskell---short for "apply." (They also use `<*>`, but who can remember that?) This works like this: ap [f] [x; y] = [f x; f y] ap (Some f) (Some x) = Some (f x) and so on. Here are the laws that any ap operation can be relied on to satisfy: ap (unit id) v = v ap (ap (ap (return compose) u) v) w = ap u (ap v w) ap (unit f) (unit x) = unit (f x) ap u (unit y) = ap (unit (fun f -> f y)) u Another general monad operation is called `join`. This is the operation that takes you from an iterated monad to a single monad. Remember when we were explaining the `bind` operation for the list monad, there was a step where we went from: [[1]; [1;2]; [1;3]; [1;2;4]] to: [1; 1; 2; 1; 3; 1; 2; 4] That is the `join` operation. All of these operations can be defined in terms of `bind` and `unit`; or alternatively, some of them can be taken as primitive and `bind` can be defined in terms of them. Here are various interdefinitions: lift f m = ap (unit f) m lift2 f m n = ap (lift f m) n = ap (ap (unit f) m) n ap m n = lift2 id m n lift f m = m >>= compose unit f lift f m = ap (unit f) m join m2 = m2 >>= id m >>= f = join (lift f m) m >> n = m >>= (fun _ -> n) m >> n = lift2 (fun _ -> id) m n Monad outlook ------------- We're going to be using monads for a number of different things in the weeks to come. The first main application will be the State monad, which will enable us to model mutation: variables whose values appear to change as the computation progresses. Later, we will study the Continuation monad. In the meantime, we'll look at several linguistic applications for monads, based on what's called the *reader monad*. The reader monad ---------------- Introduce Heim and Kratzer's "Predicate Abstraction Rule" The intensionality monad ------------------------ ... intensional function application. In Shan (2001) [Monads for natural language semantics](http://arxiv.org/abs/cs/0205026v1), Ken shows that making expressions sensitive to the world of evaluation is conceptually the same thing as making use of a *reader monad* (which we'll see again soon). This technique was beautifully re-invented by Ben-Avi and Winter (2007) in their paper [A modular approach to intensionality](http://parles.upf.es/glif/pub/sub11/individual/bena_wint.pdf), though without explicitly using monads. All of the code in the discussion below can be found here: [[intensionality-monad.ml]]. To run it, download the file, start OCaml, and say # #use "intensionality-monad.ml";; Note the extra `#` attached to the directive `use`. Here's the idea: since people can have different attitudes towards different propositions that happen to have the same truth value, we can't have sentences denoting simple truth values. If we did, then if John believed that the earth was round, it would force him to believe Fermat's last theorem holds, since both propositions are equally true. The traditional solution is to allow sentences to denote a function from worlds to truth values, what Montague called an intension. So if `s` is the type of possible worlds, we have the following situation:
Extensional types                 Intensional types       Examples
-------------------------------------------------------------------

S         s->t                    s->t                    John left
DP        s->e                    s->e                    John
VP        s->e->t                 s->(s->e)->t            left
Vt        s->e->e->t              s->(s->e)->(s->e)->t    saw
Vs        s->t->e->t              s->(s->t)->(s->e)->t    thought
This system is modeled on the way Montague arranged his grammar. There are significant simplifications: for instance, determiner phrases are thought of as corresponding to individuals rather than to generalized quantifiers. If you're curious about the initial `s`'s in the extensional types, they're there because the behavior of these expressions depends on which world they're evaluated at. If you are in a situation in which you can hold the evaluation world constant, you can further simplify the extensional types. Usually, the dependence of the extension of an expression on the evaluation world is hidden in a superscript, or built into the lexical interpretation function. The main difference between the intensional types and the extensional types is that in the intensional types, the arguments are functions from worlds to extensions: intransitive verb phrases like "left" now take intensional concepts as arguments (type s->e) rather than plain individuals (type e), and attitude verbs like "think" now take propositions (type s->t) rather than truth values (type t). The intenstional types are more complicated than the intensional types. Wouldn't it be nice to keep the complicated types to just those attitude verbs that need to worry about intensions, and keep the rest of the grammar as extensional as possible? This desire is parallel to our earlier desire to limit the concern about division by zero to the division function, and let the other functions, like addition or multiplication, ignore division-by-zero problems as much as possible. So here's what we do: In OCaml, we'll use integers to model possible worlds: type s = int;; type e = char;; type t = bool;; Characters (characters in the computational sense, i.e., letters like `'a'` and `'b'`, not Kaplanian characters) will model individuals, and OCaml booleans will serve for truth values.
type 'a intension = s -> 'a;;
let unit x (w:s) = x;;

let ann = unit 'a';;
let bill = unit 'b';;
let cam = unit 'c';;
In our monad, the intension of an extensional type `'a` is `s -> 'a`, a function from worlds to extensions. Our unit will be the constant function (an instance of the K combinator) that returns the same individual at each world. Then `ann = unit 'a'` is a rigid designator: a constant function from worlds to individuals that returns `'a'` no matter which world is used as an argument. Let's test compliance with the left identity law:
# let bind m f (w:s) = f (m w) w;;
val bind : (s -> 'a) -> ('a -> s -> 'b) -> s -> 'b = 
# bind (unit 'a') unit 1;;
- : char = 'a'
We'll assume that this and the other laws always hold. We now build up some extensional meanings: let left w x = match (w,x) with (2,'c') -> false | _ -> true;; This function says that everyone always left, except for Cam in world 2 (i.e., `left 2 'c' == false`). Then the way to evaluate an extensional sentence is to determine the extension of the verb phrase, and then apply that extension to the extension of the subject:
let extapp fn arg w = fn w (arg w);;

extapp left ann 1;;
# - : bool = true

extapp left cam 2;;
# - : bool = false
`extapp` stands for "extensional function application". So Ann left in world 1, but Cam didn't leave in world 2. A transitive predicate: let saw w x y = (w < 2) && (y < x);; extapp (extapp saw bill) ann 1;; (* true *) extapp (extapp saw bill) ann 2;; (* false *) In world 1, Ann saw Bill and Cam, and Bill saw Cam. No one saw anyone in world two. Good. Now for intensions: let intapp fn arg w = fn w arg;; The only difference between intensional application and extensional application is that we don't feed the evaluation world to the argument. (See Montague's rules of (intensional) functional application, T4 -- T10.) In other words, instead of taking an extension as an argument, Montague's predicates take a full-blown intension. But for so-called extensional predicates like "left" and "saw", the extra power is not used. We'd like to define intensional versions of these predicates that depend only on their extensional essence. Just as we used bind to define a version of addition that interacted with the option monad, we now use bind to intensionalize an extensional verb:
let lift pred w arg = bind arg (fun x w -> pred w x) w;;

intapp (lift left) ann 1;; (* true: Ann still left in world 1 *)
intapp (lift left) cam 2;; (* false: Cam still didn't leave in world 2 *)
Because `bind` unwraps the intensionality of the argument, when the lifted "left" receives an individual concept (e.g., `unit 'a'`) as argument, it's the extension of the individual concept (i.e., `'a'`) that gets fed to the basic extensional version of "left". (For those of you who know Montague's PTQ, this use of bind captures Montague's third meaning postulate.) Likewise for extensional transitive predicates like "saw":
let lift2 pred w arg1 arg2 = 
  bind arg1 (fun x -> bind arg2 (fun y w -> pred w x y)) w;;
intapp (intapp (lift2 saw) bill) ann 1;;  (* true: Ann saw Bill in world 1 *)
intapp (intapp (lift2 saw) bill) ann 2;;  (* false: No one saw anyone in world 2 *)
Crucially, an intensional predicate does not use `bind` to consume its arguments. Attitude verbs like "thought" are intensional with respect to their sentential complement, but extensional with respect to their subject (as Montague noticed, almost all verbs in English are extensional with respect to their subject; a possible exception is "appear"):
let think (w:s) (p:s->t) (x:e) = 
  match (x, p 2) with ('a', false) -> false | _ -> p w;;
Ann disbelieves any proposition that is false in world 2. Apparently, she firmly believes we're in world 2. Everyone else believes a proposition iff that proposition is true in the world of evaluation.
intapp (lift (intapp think
                     (intapp (lift left)
                             (unit 'b'))))
       (unit 'a') 
1;; (* true *)
So in world 1, Ann thinks that Bill left (because in world 2, Bill did leave). The `lift` is there because "think Bill left" is extensional wrt its subject. The important bit is that "think" takes the intension of "Bill left" as its first argument.
intapp (lift (intapp think
                     (intapp (lift left)
                             (unit 'c'))))
       (unit 'a') 
1;; (* false *)
But even in world 1, Ann doesn't believe that Cam left (even though he did: `intapp (lift left) cam 1 == true`). Ann's thoughts are hung up on what is happening in world 2, where Cam doesn't leave. *Small project*: add intersective ("red") and non-intersective adjectives ("good") to the fragment. The intersective adjectives will be extensional with respect to the nominal they combine with (using bind), and the non-intersective adjectives will take intensional arguments. Finally, note that within an intensional grammar, extensional funtion application is essentially just bind:
# let swap f x y = f y x;;
# bind cam (swap left) 2;;
- : bool = false