Support materials for "Agent-based models of strategies for the emergence and evolution of grammatical agreement" by K. Beuls and L. Steels

This web demonstration supports the following publication:

Beuls, K., & Steels, L. (2013). Agent-Based Models of Strategies for the Emergence and Evolution of Grammatical Agreement. PLoS ONE, 8(3), e58960. doi:10.1371/journal.pone.0058960.s001

An oral presentation of the paper can be obtained through the following website: digiles.vub.ac.be (login: digisteels, password: digisteels).

The experiments have been implemented using the Fluid Construction Grammar computational formalism. For an introduction to this formalism, we refer to the literature, see in particular the FCG Design patterns book and associated demonstrations. The system can be downloaded from www.fcg-net.org . All the structures shown on this website can be expanded or closed again by clicking on the boxes for linguistic units or on the syn and sem labels that open the syntactic or semantic structure respectively. Clicking on the boxes showing the parsing and production process will show the transient structure before construction application, the construction being applied, and the result of application.

We recommend using Firefox, Safari or Google Chrome for an optimal viewing of this page.

1. The Ontology and Lexicon of the Agents

Situations are generated based on an ontology, which is itself randomly generated. The ontology consists of 10 different types. Each type has a set of possible attributes. By default, the maximum number of attributes is 5. This means that one type has minimally 1 and maximally 5 attributes. A property is of the form (a-X p-X-Y Z) where X is the index of the attribute, Y that of the value and Z the object (or variable) to which the property applies. Every attribute occurs with at least two types, so that it may be a distinctive attribute in a particular situation. An example ontology is the following:

type 1
((a-3 (p-3-1 p-3-2 p-3-3) (a-4 (p-4-1 p-4-2 p-4-3)) (a-5 (p-5-1 p-5-2 p-5-3))))
type 2
((a-1 (p-1-1 p-1-2 p-1-3)) (a-3 (p-3-1 p-3-2 p-3-3)) (a-4 (p-4-1 p-4-2 p-4-3)) (a-5 (p-5-1 p-5-2 p-5-3)))
type 3
((a-2 (p-2-1 p-2-2 p-2-3)) (a-4 (p-4-1 p-4-2 p-4-3)))
type 4
((a-1 (p-1-1 p-1-2 p-1-3)) (a-4 (p-4-1 p-4-2 p-4-3)) (a-5 (p-5-1 p-5-2 p-5-3)))
type 5
((a-2 (p-2-1 p-2-2 p-2-3)) (a-3 (p-3-1 p-3-2 p-3-3)) (a-4 (p-4-1 p-4-2 p-4-3)))
type 6
((a-4 (p-4-1 p-4-2 p-4-3)))
type 7
((a-1 (p-1-1 p-1-2 p-1-3)) (a-2 (p-2-1 p-2-2 p-2-3)) (a-3 (p-3-1 p-3-2 p-3-3)) (a-4 (p-4-1 p-4-2 p-4-3)) (a-5 (p-5-1 p-5-2 p-5-3)))
type 8
((a-2 (p-2-1 p-2-2 p-2-3)) (a-3 (p-3-1 p-3-2 p-3-3)) (a-4 (p-4-1 p-4-2 p-4-3)) (a-5 (p-5-1 p-5-2 p-5-3)))
type 9
((a-1 (p-1-1 p-1-2 p-1-3)) (a-3 (p-3-1 p-3-2 p-3-3)) (a-4 (p-4-1 p-4-2 p-4-3)))
type 10
((a-1 (p-1-1 p-1-2 p-1-3)) (a-2 (p-2-1 p-2-2 p-2-3)) (a-4 (p-4-1 p-4-2 p-4-3)) (a-5 (p-5-1 p-5-2 p-5-3)))

The lexicon is generated automatically based on the types in the ontology. It is shared by all agents in the population and also new-born agents receive a copy of it. An example of a lexical entry in the FCG implementation is as follows (You can click on the box and then on sem and syn in order to see the internals.) The construction associates the property (a-1 p-1-2) with the string shuqfon

shuqfon-cxn

The lexicon contains on average 286 words: 15 of them have a single predicate meaning (e.g. (p-1-3)), 63 have meanings with two predicates and 208 words have the maximum size meaning.

the lexicon

2. Formal Marker Strategy (cf. Figure 3)

First the speaker produces an utterance. Two grammatical constructions (expand-word-group and create-word-group) group all words that refer to the same object. Once an utterance is produced the speaker re-enters the utterance to detect whether there is any ambiguity in interpretation.

Current situation:
((a-1 p-1-2 o-1) (a-2 p-2-2 o-1) (a-3 p-3-1 o-1) (a-1 p-1-2 o-2) (a-2 p-2-3 o-2) (a-3 p-3-2 o-2) (a-1 p-1-3 o-3) (a-2 p-2-2 o-3) (a-3 p-3-1 o-3))
The lexicon (speaker and hearer):
Word group constructions:
Speaker Topic:
o-1 + o-2
Production process:
Transient structure:
Utterance:
("shuqfon" "iqvu" "sizhic")
Re-entrance parsing:
Analysis:
Parsed meaning
((a-3 p-3-2 ?object-181048) (a-1 p-1-2 ?object-181046) (a-2 p-2-2 ?object-181047))
Possible topics
o-1 + o-2
o-1 + o-2 + o-3
o-2 + o-3

The speaker detects more than one possible interpretation and triggers a repair strategy which consists in creating new formal markers. The speaker then reproduces the utterance with these new constructions and performs re-entrance again to test whether the utterance is now without ambiguity.

New markers:
Re-production:
New utterance:
("iqvu" "-ta" "shuqfon" "-ta" "sizhic" "-ti")
Re-parsing:
Result:

Now the hearer parses the utterance. Because there are unknown markers, he triggers a repair strategy that will create constructions for the two markers.

Input utterance:
("iqvu" "-ta" "shuqfon" "-ta" "sizhic" "-ti")
Hearer parsing:
Result:
New markers:
Re-parse
Result:
Meaning:
((a-1 p-1-2 ?object-181048) (a-3 p-3-2 ?object-181048) (a-2 p-2-2 ?object-181047))
Possible topics:
o-1 + o-2

Notice how the variables ?object-68671 and ?object-68669 have become both equal to ?object-68671 so that there is no more ambiguity about which words belong together.

3. Strategy 2A: Meaningful markers - no reuse (Figure 6)

We now show what the meaningful marker constructions look like using a concrete example.

Current situation:
((a-5 p-5-3 o-25) (a-1 p-1-3 o-25) (a-3 p-3-3 o-25) (a-4 p-4-3 o-25) (a-5 p-5-2 o-26) (a-1 p-1-2 o-26) (a-2 p-2-1 o-26) (a-3 p-3-3 o-26) (a-4 p-4-2 o-26) (a-5 p-5-1 o-27) (a-1 p-1-1 o-27) (a-2 p-2-1 o-27) (a-3 p-3-2 o-27))
The initial grammar:
Speaker topic:
(o-28 o-29)
Production process:
Utterance
("pojcog" "opjae" "aerlae" "quznir" "dodvaex")

Because there is more than one interpretation possible, the speaker invents new markers. The internals of these constructions can be seen by clicking on sem or syn or on the individual boxes. The construction constrains the semantic features of the word of which it can be a suffix. The suffix is a syntactic subunit of the word to which it is attached.

New markers:
Re-Production:
Utterance:
("tatjoz" "-qet" "puwpuid" "-qet" "gixchaeh" "-zih" "jeylon" "-zih")

4. Strategy 2B: Meaningful markers - with lexical reuse (Figure 8

The implementation is similar as with Strategy 2B, except that the form and semantic contraints of the marker are derived from an existing word in the lexicon.

Here is an example of a meaningful marker based on reuse:
Production process:
Transient structure:
Utterance:
("lichnaq" "-gixchaeh" "mefyesh" "-gixchaeh" "gixchaeh" "-gixchaeh" "nuuli" "-qauri" "tatjoz" "-qauri" "opjae" "-qauri")

5. Strategy 4: Recategorization

In the following example the feature matrices have two dimensions (a-1 and a-2) but only 2 values, so that other values for these attributes are undetermined and markers cannot apply.

Example of marker M2:
Example of controller noun saeshkuip:

The marker cannot be used with this noun. The noun has variables in its feature matrices (click on sem and syn to see the internals) meaning that it is so far under determined. The noun can be coerced by changing these variables into the concrete values of the marker, so that a complete parse becomes possible.

After coercion:
nil
Utterance:
("fihbaesh" "ta" "fojveg" "ta" "izche" "ta" "shouyo" "ta" "uitmui" "tu" "sodvac" "tu" "shouyo" "tu" "aywa" "tu")
Parsing:
Result: