Dependent types and spatial relations in pictures

Important: This is a copy of the original code from Github.

Before we start, this is a literate Agda file. That means you can either read it online or directly load and evaluate it in Emacs using the agda-mode (assuming that you have Agda readily set up).

This project uses dependent types to express spatial constraints on objects in a picture (which we call a “scene”). The objects of a scene type can be translated in either the pictures as graphics or picture descriptions as natural language. We attempt to require little previous knowledge about dependent types and other topics. But some understanding of (functional) programming can be very herlpful. The original implementation has been done in Grammatical Framework (GF) and later been translated into Agda. Even though plenty of inspiration has come from using GF, we try our best to make this document understandable without any knowledge of GF.

We start this project by defining an “abstract” module. It defines the data types which we need to describe our pictures. Later we can add modules to translate these data types into other representations. We call this module “SpatialAbs”

module SpatialAbs where


Dependent types are great to model constraints. To put it very shortly, a dependent type is a type that depends on values of another type. A simple example is the definition of vectors, i.e. lists with associated length. The type of a vector depends on its length. A big issue with lists in computer science is, that the e.g. the basic operation of getting the first element (head) of a list does not work on empty lists. It is possible to catch this as an error, but it would be nicer to not allow this in the first place. Dependent types can solve this issue because the head function is only defined for vectors of length larger than 0 and this constraint can be expressed on the type level. We will now have a look on how vectors can be implemented in Agda:

open import Agda.Builtin.Nat

data Vector (A : Set) : Nat → Set where
  []  : Vector A zero
  _∷_ : {n : Nat} → A → Vector A n → Vector A (suc n)

We use the built-in definition of natural numbers, which is essentially the same as the numbers we will use later, based on the Peano axioms. We will see the definition of these numbers later. For the moment zero is a natural number representing 0 (what a surprise) and we can form any natural number n by sequence of adding 1 n times to 0, i.e. 1+1+…+0. Addind 1 is called the successor (suc).

Because we want to have vectors working for any type, we use a type variable (A : Set) in the definition (Set is the type of types in Agda). And the part Nat → Set defines a vector as a dependent type, depending on a number of type Nat, the length of the vector.

To form an object of a type we use a constructor for this type. A constructor is a function that can take arguments and produces an object of the type. For vectors we have twoc constructors. Vector are similar to a list with [] being the empty vector and using :: (spoken as cons) to add a value to the vector. Both are type constructors for vectors. The constructor for the empty vector sets the length in the type (!) to 0. The constructor :: takes several arguments: An implicit number n (in curly brackets), a value of type A and a vector of length n containing elements of type A. As a result it again produces a vector containing values of type A but with the type containing the length plus 1. Implicit arguments are arguments that don’t have to be given by the user but instead can he infered.

Now we can define a safe version of head only working on vectors of length greater than 0 (expressed as suc n, which is always at least 1 greater than zero):

head : {n : Nat} → {A : Set } → Vector A (suc n) → A
head (x ∷ v) = x

The function type of this function guarantees that we cannot try to get the head of an empty list, because an empty list does not match the argument type of the function. We again have implicit arguments for a number and a type (the type of elements in the vector).

Let’s see how we can use this type and function. We define two vectors, one empty and one with three elements called testVector. The non-empty vector can be used with our head function, the other can not. The reason is connected to the implicit arguments, specifically the first argument. It is a natural number that, when 1 added to it is the length of the vector. For our test vector we know it has to be 2 and the type of the elements is Nat, so we can give them explicitly. But it also works when skipping these parameters (in the case of h'). Agda can magically infer the value from the definition of the types (the magic of type inference!). But for the empty vector Agda fails. There is no such natural number, that when 1 is added it is 0. For that reason we cannot use h''.

emptyVector : Vector Nat 0
emptyVector = []
testVector : Vector Nat 3
testVector = 1 ∷ (2 ∷ (3 ∷ []))

h : Nat
h = head {2} {Nat} testVector
h' : Nat
h' = head testVector
-- h'' = head {!!} {Nat} emptyVector

But dependent types can also be very useful when modelling concepts closely related to natural language, especially concepts of semantics. This document explains how dependent types can be used to express spatial relationship between objects in a picture and how the same type-theoretic representation can be translated into both graphical pictures and natural language descriptions.

Getting started

Before we can dive into more complex topics, we have to define a concept of numbers. An easy way to define natural numbers is using the Peano axioms. We could use the built-in numbers as we did with the vectors before, but it cannot hurt to repeat the definition here:

data Num : Set where
  z : Num
  s : Num → Num

This definition basically says, that zero (z) is a natural number (Num) and for each number, the succesor (s) is also a number. We need this type later to place objects and check if a position is valid for a certain spatial relationship.

To place object we use a grid with x and y directions, ranging from 0 to a maximum for each direction. The grid has the following shape:

|     |     |     |     |     |
| 0,2 | 1,2 | 2,2 | 3,2 | 4,2 |
|     |     |     |     |     |
|     |     |     |     |     |
| 0,1 | 1,1 | 2,1 | 3,1 | 4,1 |
|     |     |     |     |     |
|     |     |     |     |     |
| 0,0 | 1,0 | 2,0 | 3,0 | 4,0 |
|     |     |     |     |     |

To avoid some tedious typing, we can define a few number constants for the numbers from 1 to 5 and for the maximum x and y valus.

n1 : Num
n1 = s z

n2 : Num
n2 =  s n1

n3 : Num
n3 = s n2

n4 : Num
n4 = s n3

n5 : Num
n5 = s n4

maxx : Num
maxx = n5

maxy : Num
maxy = n4

After defining our numbers we can move on to the description of our pictures. We already mentioned a grid on which objects can be placed. The next step is to define these basic objects. The type of objects is simply defined by listing all possible values. All types we have defined so far are simple types. But they will be later used in the definition of the dependent types describing, e.g. what objects can appear in which position.

data Object : Set where
  otree : Object
  ohouse : Object
  osun : Object
  operson : Object
  otable : Object
  obox : Object
  oball : Object

Another simple type is the list of all valid relations between objects. This type is again defined listing all values. We later have to define what exactly it means to be beside, in or above of another object.

data Relation : Set where
  rbeside : Relation
  rleftof : Relation
  rrightof : Relation
  rin : Relation
  rabove : Relation
  rontopof : Relation
  rnextto : Relation

But before moving on, we need to add a bit to our numbers. To be able to use our numbers in a meaningful way, we need to know when two numbers are equal or when one number is smaller than another. We can define these two relations using dependent types for the first time.

We define two new data types, modeling if two numbers are equal or if a number is smaller than another one. These are not predicates in the traditional sense, they are data types, but if we can construct a value of such a type, it counts as a proof that the relation holds.

Equality is defined by reflexivity, i.e. a number is equal to itself, and by symmetry, i.e. if x=y then also y=x. The second case is probably not necessary but there are some issues with dependent types in GF which can be circumvented by including this case. Going back to what was just said. Given any number, we can construct a proof object that it is equal to itself. And for each two numbers if we know that x=y then we can also prove that y=x, i.e. from the proof object IsEqual x y we can construct a proof for IsEqual y x.

data IsEqual : Num → Num → Set where
  equal : (n : Num) → IsEqual n n
  sequal : (n n' : Num) → IsEqual n n' → IsEqual n' n

The less-than relation is defined the following way: - zero is less than the successor of every other number (lessz) - If n1<n2 then also the succesor of n1 is smaller than the succesor of n2 (lesss) - by transitivity, if n1<n2 and n2<n2 we also know that n1<n3 (lesst)

data IsLess : Num → Num → Set where
  lessz : (n : Num) → IsLess z (s n)
  lesss : (n1 n2 : Num) → IsLess n1 n2 → IsLess (s n1) (s n2)
  lesst : (n1 n2 n3 : Num ) → IsLess n1 n2 → IsLess n2 n3 → IsLess n1 n3

All these number parameters involved could be turned into implicit arguments, i.e. they don’t have to be given explicitly and don’t show up in the proofs (or functions). But because we want to build the same terms as we would get from GF we include them es explicit arguments.

These two relations come in handy later, when actually placing objects. But before placing objects, we can clasify them. Not every object can take every position. For example it is quite difficult to place something on top of a ball. Of course we could define for each object separately, where it can be used. But it is easier to define classes of objects that can be used in the rules about how objects can be placed. Some of these classes are only defined for one type of object, so we could use this object directly. But having these classes, we can add objects later more easily.

data BaseObject : Object → Set where
  tablebase : BaseObject otable
data OutsideObject : Object → Set where
  houseout : OutsideObject ohouse
  boxout : OutsideObject obox
data InsideObject : Object → Set where
  tablein : InsideObject otable
  boxin : InsideObject obox
  ballin : InsideObject oball
data BesideObject : Object → Set where
  treebeside : BesideObject otree
  housebeside : BesideObject ohouse
  personbeside : BesideObject operson
  tablebeside : BesideObject otable
  boxbeside : BesideObject obox
  ballbeside : BesideObject oball
data OnTopObject : Object → Set where
  ballontop : OnTopObject oball
  boxontop : OnTopObject obox
data AboveObject : Object → Set where
  sunabove : AboveObject osun
  ballabove : AboveObject oball
data BelowObject : Object → Set where
  treebelow : BelowObject otree
  housebelow : BelowObject ohouse
  personbelow : BelowObject operson

Now that we have defined the objects, spatial relations and classes of objects, we can define a new dependent type describing what relation is valid between which two objects.

The first four just rely on the classes we defined. For example, if both objects are in the class BesideObject they can be combined in the Beside relation.

But then some of the relations are related. To express this, we have a function that tells us, that if something has type A, it is also of type B. Such a function is often called a a type coercion. The problem we try to solve is, that we only defined that objects can be besides each other, but our relations are more fine-grained. Objects can be either to the left or to the right or directly next to or further away, but all of them qualify as “besides”. Of course we could say, that every BesidesObject is also a NextToObject and then define a validnextto constructor for ValidRel. This would mean that we have a type coercion defined on the different kind of objects. A shorter solution is given below, where we define, that every pair of objects, that can be besides each other, can also be next to each other and it does not matter if it is to the left or to the right, to be a valid relation. This solution probably does not really qualify as a coercion but solves a similar problem.

Finally, we handle a few special cases, where we don’t want to use the general classes but only handle a few objects separately. It is not nice to put a person into a box, so we did not define Person as an InsideObject, but people can be in houses.

data ValidRel : Relation → Object → Object → Set where
  validbeside : {o1 o2 : Object} → BesideObject o1 → BesideObject o2 → ValidRel rbeside o1 o2
  validin : {o1 o2 : Object} → InsideObject o1 → OutsideObject o2 → ValidRel rin o1 o2
  validabove : {o1 o2 : Object} → AboveObject o1 → BelowObject o2 → ValidRel rabove o1 o2
  validontop : {o1 o2 : Object} → OnTopObject o1 → BaseObject o2 → ValidRel rontopof o1 o2
  -- Coerce relations
  nexttoisbeside : {o1 o2 : Object} -> ValidRel rbeside o1 o2 -> ValidRel rnextto o1 o2
  leftofisbeside : {o1 o2 : Object} -> ValidRel rbeside o1 o2 -> ValidRel rleftof o1 o2
  rightofisbeside : {o1 o2 : Object} -> ValidRel rbeside o1 o2 -> ValidRel rrightof o1 o2
  ontopofisabove : {o1 o2 : Object} -> ValidRel rontopof o1 o2 -> ValidRel rabove o2 o1
  -- Special cases
  boxinhouse : ValidRel rin obox ohouse
  personinhouse : ValidRel rin operson ohouse
  treeinbox : ValidRel rin otree obox

Now we know which objects can be combined with which relation, but we still need to paint them on our canvas, or at least place them on our grid. We have two objects and each of them has x and y coordinates. The next type makes sure that the two positions are actually on the grid, using the variables maxx and maxy we defined above. That already makes sure that we don’t place anything outside the grid. Here we extensively use our IsLess type for the upper limit and have 0 as the lower limit because that is the smallest number we can express.

data InRange : Num -> Num -> Num -> Num -> Set where
  -- Restrictions on positions
  inrange : {x1 y1 x2 y2 : Num} -> IsLess x1 maxx -> IsLess x2 maxx -> IsLess y1 maxy -> IsLess y2 maxy -> InRange x1 y1 x2 y2

The next step is to also rule out all placements that don’t match the meaning of the relations. So we have to define what it means, if something is “in” something else, “next to” something and so on. So far we settled on the following constraints: - “in” means that the objects are in the same place, i.e. the x and y coordinates of the two objects are equal. Additionally, the objects have to be placed on the floor, i.e. on y coordinate 0. - “left of” means that both objects are on the floor and the x coordinate of the first object is smaller than the one of the second object. This is potentially more restrictive than necessary because it could be acceptable that something is left of something else even if it is above it. - “right of” is similar to “left of”, except that the x coordinate of the second object is smaller than the one of the first object - “next to” is split in the two cases, left of or right of. The difference to the two previous cases is, that we want to place the objects next to each other, i.e. the x coordinate should be exactly one more or less, depending if it should go to the left or the right - For “above” we want to have the same x coordinate but the first object has a larger y coordinate than the first. And of course the second object is on the floor again. - Finally, for “on top of”, we want to have the same as above, but the y coordinate should be exactly one larger than the other

These constraints are pretty much ad-hoc, but they showcase how spatial relations can be translated into coordinates.

data ValidPos : Relation -> Num -> Num -> Num -> Num -> Set where
  validinpos : {x1 y1 x2 y2 : Num} -> IsEqual x1 x2 -> IsEqual y1 y2 -> IsEqual y2 z -> InRange x1 y1 x2 y2 -> ValidPos rin x1 y1 x2 y2
  validleftofpos : {x1 y1 x2 y2 : Num} -> IsEqual y1 y2 -> IsEqual y1 z -> IsLess x1 x2 -> InRange x1 y1 x2 y2 -> ValidPos rleftof x1 y1 x2 y2
  validrightofpos : {x1 y1 x2 y2 : Num} -> IsEqual y1 y2 -> IsEqual y1 z -> IsLess x2 x1 -> InRange x1 y1 x2 y2 -> ValidPos rrightof x1 y1 x2 y2
  validnexttoleftpos : {x1 y1 x2 y2 : Num} -> IsEqual y1 y2 -> IsEqual y1 z -> IsEqual x2 (s x1) -> InRange x1 y1 x2 y2 -> ValidPos rnextto x1 y1 x2 y2
  validnexttorightpos : {x1 y1 x2 y2 : Num} -> IsEqual y1 y2 -> IsEqual y1 z -> IsEqual (s x2) x1 -> InRange x1 y1 x2 y2 -> ValidPos rnextto x1 y1 x2 y2
  validabovepos : {x1 y1 x2 y2 : Num} -> IsLess y2 y1 -> IsEqual y2 z -> IsEqual x1 x2 -> InRange x1 y1 x2 y2 -> ValidPos rabove x1 y1 x2 y2
  validontopofpos : {x1 y1 x2 y2 : Num} -> IsEqual (s y2) y1 -> IsEqual y2 z -> IsEqual x1 x2 -> InRange x1 y1 x2 y2 -> ValidPos rontopof x1 y1 x2 y2

Now we have all the parts and we can define how we can build a picture or a Scene from it. We define a scene as a new simple type, but to construct it we need quite a bit of stuff: As a start the two objects, their coordinates and the spatial relation. But then we also have to make sure that the objects can be placed using this relation (ValidRel) and the positions work for the relation (ValidPos). Only if we can give a proof for these properties, we can create a complete picture.

data Scene : Set where
  constraintPlace : (o1 o2 : Object) -> (x1 y1 x2 y2 : Num) -> (r : Relation) -> ValidRel r o1 o2 -> ValidPos r x1 y1 x2 y2 -> Scene

Finally, we can give an example. It is a person in a house and both objects are at the coordinate (3,0). The relation is “in” and we know that we can put a person into a house because of our constructor personinhouse. The messy part is the proof objects that guarantee that the objects can be placed in the way to each other that they are actually inside each other and that they are really placed on the grid. Translated into a natural language description we would simply get the sentence “the person is in the house”.

example : Scene
-- example = constraintPlace operson ohouse n3 z n3 z rin personinhouse (validinpos (s (s (s z))) z (s (s (s z))) z  (equal (s (s (s z))))  (equal z) (equal z)  (inrange (s (s (s z))) z (s (s (s z))) z (lesss (s (s z)) (s (s (s (s z)))) (lesss (s z) (s (s (s z))) (lesss z (s (s z)) (lessz (s z))))) (lesss (s (s z)) (s (s (s (s z)))) (lesss (s z) (s (s (s z))) (lesss z (s (s z)) (lessz (s z))))) (lessz (s (s (s z)))) (lessz (s (s (s z))))))
example = constraintPlace operson ohouse n3 z n3 z rin personinhouse (validinpos (equal (s (s (s z)))) (equal z) (equal z) (inrange (lesss (s (s z)) (s (s (s (s z))))
                                                                                                                                       (lesss (s z) (s (s (s z))) (lesss z (s (s z)) (lessz (s z))))) (lesss (s (s z)) (s (s (s (s z))))
                                                                                                                                                                                                         (lesss (s z) (s (s (s z))) (lesss z (s (s z)) (lessz (s z))))) (lessz (s (s (s z)))) (lessz (s (s (s z))))))

As an additional experiment, we can also use Agda to translate the data types we defined here into natural languages. A simple example is the translation to English and a more challenging example is the translation to German. To translate into pictures we implemented an Agda module that generates HTML output that shows the pictures.