Computational Commonalities Between Language and Space - Petnica - August 2014

Common aspects of cognitive computations in language and in

spatial cognition

Boban Arsenijevi,

University of Ni

1

Similar conditions

Most investigations of cognitive differences between humans and other animals look for the counterparts of human capacities (language, arithmetic, music) in animals.

2

Goal of the talk

In this talk, I present an (interrupted) research which:

1. considers the cognitive apparatus as an automaton (i.e. a mathematical model, a formal grammar) and

2. looks for evidence of the degree of computational complexity of these automata in humans and other animals,

3. by looking at the spatial cognition, on the background of some well known properties of language.

3

Spatial cognition

Probably the earliest cognitive capacity developed.

Universally available in all animal species.

Very sophisticated already in relatively primitive animals / brains (e.g. insects).

Immediacy of exposure: organisms take up space, all sensory inputs relate to space.

Includes information from multiple sensory sources and cognitive domains.

4

Language

A very recent evolutionary development.

Capacity of the cognitive apparatus of only one species.

Exposure conditioned by socio-cultural conditions.

Includes information from multiple sensory sources and cognitive domains.

5

Language and space

In all languages, at least 10, often more than 20 words with spatial semantics among the 50 most frequent words in the language (mostly prepositions and adverbs, and some nouns).

Only a limited subset of the possible spatial relations finds expression in language.

Still, the semantic differences between spatial expressions are often subtle and very complex (beneath : below : under; S-C pred : ispred).

6

Context representation

Both language and spatial cognition heavily rely on a dynamic representation of the relevant context.

In spatial cognition, the context representation is referred to as the cognitive map.

In language, the corresponding notion is the discourse representation.

7

Modeling spatial cognition

Core elements: cognitive maps and the mechanisms of their update and retrieval.

Cognitive maps are based on an ontology of four basic members:

1. landmarks,

2. geometrical structures,

3. locations and

4. properties of locations.

8

Example

Location: palm

Properties: shade, fruit, snakes. Landmark:

mountain peak

Geometrical structure: straight line

9

Similarities functional perspective

Both representations are verified against a sensory input: relied upon in acts of motion / communication, and updated at mismatches.

Both representations are plausibly modeled as webs of entities, characterized by their properties and connected by their mutual relations.

The discourse usually contains spatial information, and cognitive maps are often updated by linguistically transmitted information.

10

Processing similarities

General architecture: information flow between the sensory input, the context representation and the motorics.

Sensory

input

Motoric

system

Processing

Context

representation

11

Similarities neurocognitive perspective

A central role of hippocampus storing, updating and retrieving the information in the context representation.

Both prefrontal zones episodic memory, semantic contents.

Right (superior and) inferior parietal gyrus more prominently involved in spatial cognition, left superior and) inferior parietal gyrus in language.

12

Interactions: language and navigation types

Languages (and cultures) with prominent alocentric, i.e. egocentric, relations in expressing spatial relations (between the left/right type of relations and the north/south, central/peripheral, uphill/downhill type of relations).

Controversial data about the influence of these parameters on the navigation preferences and capacities (Gleitman and Li).

13

Deeper, computational similarities?

Grammar involves recursive computations producing hierarchical structures.

my friends

little brothers

favorite toy

Are there such structures in spatial cognition?

14

Recursive self-embedding in grammar

Some of the explanations for recursive structures in language: the output of one cycle of computations may become the input of the next cycle.

Processng Context

representat.

x [[x]]

f([[x]])

15

Examples

[[[[my friends] friends] friends] friend]

[the rabbit chasing [the rabbit chasing [the rabbit chasing [the rabbit...]]]]

[you said [that you said [that you said ]]]

[a box in [a box in [a box in [a box ]]]]

16

Recursive computations and cognition

H1: the capacity for recursive computations is the main differential characteristic of human cognition in general (Hauser et al. 2002).

Prediction: recursive spatial computations in humans, but not in other animals.

H2: recursive computations differentiate the language capacity and its derivates (arithmetic? node-tying?) from all other cognitive capacities.

Prediction: no recursive structures in spatial cognition in any species, including humans.

17

Recursion in spatial cognition

Similarities outlined between language and spatial cognition, especially in the processing-memory architecture, imply such a possibility.

Can we identify clear instances of recursive computations in humans?

Many animal species have a spatial cognition more sophisticated than that in humans does it offer evidence for recursive modeling?

18

Recursive route-planning experiment

19

(Arsenijevi et al. unfinished)

Recursive computations

Features of the Location1: {path_from_here, distance_from_here, water}

Features of the Location2: {path_from_here, distance_from_here, path_to_L1, food}

[L2 path_from_here, distance_from_here, path_to_L1, food [path_from_here, distance_from_here, water]]. (path_to_L1 selects for L1 as a complement)

Recursive rule: pathY X route-embed Y in X

20

Recursive licensing of paths

21


L1

[path_from_here,

distance_from_here,

water]

L2

[path_from_here,

distance_from_here,

path_to_L1,

food]

Graphical tree representation

22

The same type of recursive computations as in grammar, whether top-down or bottom-up.

{here, pathL1, pathL2}

{pathhere, distancehere, pathL1, food}

pathhere, distancehere, water

Parallels in language

23

The same type of recursive computations as in grammar, e.g. relative clauses.

I bought the booki

whichi you bough in the shopj

thatj we visited yesterday

Recursive computations in blocking

The blocking / sub-prioritizing of the other path also requires recursive computations:

[Here here, pathL1, pathL2 [L1 pathhere, distancehere, water *[L2 pathhere, distancehere, pathL1, food]]]. (absence of path_to_L2 in L1 blocks L2 as a complement)

The negative evaluation of the route is determined by the crashed maximal derivation, i.e. by the fact that the maximal converging derivation is smaller

24

Recursive computing for blocking

25


L1

[path_from_here,

distance_from_here,

water]

L2

Graphical tree representation

26

The same type of recursive computations as in grammar, whether top-down or bottom-up.

{here, pathL1, pathL2}

{pathhere, distancehere, water}

pathhere, distancehere, pathL1, food

Parallels in language

27

A linguistic parallel would be e.g. a non-relative clause, which opposes embedding.

I bought the booki

whichi you bough in the shop

I visited you yesterday

Conclusion

If, upon the completion of the experiment, the trend observed in the initial phase is confirmed, then at least some animals have the recursive computational capacity in the spatial cognitive domain.

This will falsify both H1 and H2 as formulated above: recursive computations are neither exclusively human nor exclusively linguistic.

28

Thank you!

29