Recent press reports talked about a language that has only 3 verbs. Researchers claim that this language (Jingulu) may form a basis for creating a language that leads to better communication between humans and artificial intelligence systems.

The basis for these reports is a paper in the journal Frontiers in Physics: JSwarm: A Jingulu-Inspired Human-AI-Teaming Language for Context-Aware Swarm Guidance, by Hussein A. Abbass, Eleni Petraki and Robert Hunjet (2022) https://www.frontiersin.org/articles/10.3389/fphy.2022.944064/full  

I discuss below Jingulu and the language (JSwarm) developed by the paper’s authors reflecting insights form Jingulu.

Jingulu

Jingulu is a language spoken by the Jingili people around the town of Elliott in Australia’s Northern Territory. There is a detailed description of the language in A grammar of Jingulu, an Aboriginal language of the Northern Territory, Rob Pensalfini, 2003, https://openresearch-repository.anu.edu.au/bitstream/1885/146163/1/PL-536.pdf

Pensalfini describes the language as critically endangered—in fact, moribund. He says that there are only some 12 to 15 competent speakers of Jingulu and perhaps another 50 who speak it to some degree, though nobody speaks the language regularly today.

Verbs in Jingulu

Jingulu verbal structures have the following 3 components:

coverbal root – inflection – light verb

Coverbal roots can only appear:

• prefixed to a light verb;
• prefixed to a suffix that turns them into nominals or adverbs; or
• in some subordinate clauses interpreted as having the same verbal features as the main clause.

The inflection indicates the (grammatical) person and number of the subject and of the object.

Light verbs are verbs that are ‘semantically bleached’, containing little inherent meaning of their own. English examples of light verbs include take (in take a look) and have (in have a go): in those verbal phrases, the nouns look and go contribute almost all of the meaning.

The light verbs are bound morphemes;

• they cannot appear separately and are always be preceded by the inflection— and in many case also by a coverbal root.
• phonologically, they are suffixes on the inflection (and on the coverbal root, if there is one).

Example of verbs in Jingulu

Abbass, Petraki and Hunjet provide the following 3 examples (I have picked out the inflection and the light verb in bold):

(1) Kirlikirlika darra-ardi jimi-rna urrbuja-ni.
galah eat-go that(n)-FOC galah_grass-FOC
‘Galahs eat this grass.’

(2) Aja(-rni) ngaba-nya–jiyimi nginirniki(-rni)?
what(-FOC) have-2sg-come this(n)(-FOC)
‘What’s this you’re bringing.’

(3) Ngaja-mana–ju.
see-3MsglO-do
‘He is looking at us.’

(In the examples, FOC represents contrastive focus, 3Msg represents 3rd person masculine subject and IO represents 1^st person object)

Abbass, Petraki and Hunjet comment that:

• in example 1, eating the grass indicates that the subject needs to move away from the subject’s current location by ‘going to the grass. I
• in example 2, questioning what a person brings depicts something coming from the subject’s current location to our location.
• in example 3, ‘looking at us’ does not require any movements.

Light verbs in Jingulu

Unlike English, Jingulu has only 3 light verbs:

• One expresses motion away from the speaker or from some other salient referent. If used without a coverbal root, it translates into English as ‘go’.
• Another expresses motion toward the speaker or toward some other salient referent. If used without a coverbal root, it translates into English as ‘come’.
• The 3^rd one does not specify motion, but does not necessarily imply lack of motion. If used without a coverbal root, it translates into English as ‘do’ or ‘be’.

The following table shows the forms of the 3 light verbs:

Notes to the table:

1. The table shows only the main forms, not variants that appear in some contexts.
2. The light verb meaning ‘do’, ‘be’ also has forms for a distant past (/-marri/), habitual (/-ardi/) and past habitual (/-ka/).

As the table shows, Jingulu light verbs do not inflect for tense. Completely separate forms exist for different tenses. English too uses this process (known as suppletion) for a few verbs (eg went is the past form corresponding to go.)

JSwarm

The authors of the paper have developed a language called JSwarm, intended for use in interactions between humans and decentralised ‘swarms’ of artificial intelligence ‘agents’ (‘AI agents’).

Example: sheepdog problem

What this means can be illustrated by a shepherding problem, inspired by sheepdogs mustering sheep. The shepherding (teaming) system consists of:

• a swarm (analogous to sheep) to be guided
• an actuator agent (analogous to a sheepdog’s body) with the capacity to influence the swarm
• an AI-shepherd (analogous to sheepdog cognition) with the capacity to autonomously guide the actuator agent (sheepdog body) to achieve a mission and execute human intent;
• a human team (analogous to farmers) with the intention of moving the swarm. The human team interacts with the AI-shepherd, monitors, understands, and commands it, and takes corrective action when the AI-shepherd deviates from the human team’s intent.

Using JSwarm in the Sheepdog problem

JSwarm would be used:

• within the swarm (sheep) and actuator-agent, describing their actions and behaviour;
• in interactions between the actuator agent (sheepdog body) and AI-shepherd (sheepdog cognition)
• in interactions between the AI-shepherd (sheepdog cognition) and human team (farmers)

To be useful in all these contexts, JSwarm is designed to be used easily by agents with greatly differing cognitive capacities. In this example:

• sheep are cognitively the simplest agents with simple survival goals.
• sheepdogs have more complex cognition because they must execute a farmer’s intention autonomously.
• the cognition of farmers/shepherds is even more complex because they must understand the sheepdogs’ capabilities and issue commands to them.

Examples of possible real-world applications

The paper presents the sheepdog problem as a simple illustration of how their model works. The authors give some other examples of possible real-world applications:

• a medical practitioner controlling an external robot (sheepdog cognition) that controls a chemical substance (sheepdog body) that a swarm of nano–robots (sheep) reacts to.
• a swarm of under-water vehicles cleaning the ocean.
• a swarm of uncrewed aerial vehicles surveying a large area for the mining industry.

Structure of JSwarm sentences

The authors based the structure of Jswarm on one feature of Jingulu: its separation between the 3 light verbs and coverbal roots. (Unlike Jingulu, Jswarm has no inflection.) Sentences in JSwarm have the following structure. The verbal part (shown below in bold is reminiscent of Jingulu verbal structures).

Subject.Light verb [go / come /do].Supporting verb.obj

JSwarm has three types of sentences, discussed below:

1. behaviour sentences used in regulating behaviour and actions
2. intent sentences that send commands and goals to the AI
3. state-information sentences that send information about the state the AI is in

Behaviour sentences

The behavioural sentences are used by the swarm (sheep) and by the actuator agents (sheepdog bodies) in carrying out actions. These sentences have the following structure:

Behaviour: Subject.Verb(Space.TimeDelay.TimeDuration).SupportingVerb(Space.TimeDelay.TimeDuration).Object

Here are 3 examples of behaviour sentences:

• Dog1.Come((x = 40,y = 40).0.30).Collecting(NULL.0.20). Sheep5
  ‘Within 30 seconds from now, dog Dog1 needs to start moving to arrive at location (40,40) and spend 20 seconds collecting sheep Sheep5.’
• Sheep4.Go((x = 50,y = 70).0.50).Escape(NULL.0.30).Dog1
  ‘Now (0 delays), Sheep Sheep4 needs to go to location (50,70) within 50 seconds and escape dog Dog1 for 30 seconds.’
• Dog1.Do(NULL.10.0).Sitting(NULL.0.30)
  ‘Dog Dog1 needs to wait for 10 seconds then sit in its location for 30 seconds.’

Intent sentences

The intent statements send commands and goals to the AI-shepherd. These statements have the following structure:

Subject.DO.SupportingVerb.Object

Here is an example of an intent sentence:

• Dog1.Do.Herd.Herd1
  ‘Dog1 needs to herd group Herd1.’

State-information sentences

The state-information sentences tell humans what state the AI-shepherd is in.

They have the following structure:

States: Subject.DO.SupportingVerb(Space.TimeDuration).Object

Here is an example of a state information sentence:

• Dog1.Do.Collecting(180).Sheep5
  ‘Dog Dog1 is collecting sheep Sheep5 for 180 seconds’

What is so good about Jswarm?

Jingulu separates verbal information about tense and direction (on the light verb) from content (on the coverbal root). The authors claim that this separation makes Jingulu an ideal natural language for representing spatial movements between entities and the exchange of communication messages, including commands, among the agents. They claim that this separation would also be useful in an artificial language for communication between humans and swarms of artificial intelligence agents.

I couldn’t see from the paper why the authors think Jswarm would work so well as an interface between people and machines. The reasons seem to include the following, but I couldn’t follow all the logic:

• Jswarm separates syntax (the 3 light verbs) from semantics (the coverbal roots). The paper says this separation allows the syntax to remain intact when it is applied in different domains with different semantic layers.
• The 3 light verbs deal with all aspects of space. Space can be conceived broadly in both physical terms (as in the sheepdog example) and in more abstract terms—for example, in the space of ideas, an idea might come to a person or a person might go to an idea.
• Having only 3 light verbs makes the syntactic component simple and unambiguous.   
• A structure with only three main verbs: do, go and come ‘is most efficient for communication in attraction-repulsion equations-based distributed AI-enabled swarm systems’. The paper didn’t explain to me why this is.
• Natural human languages are too rich and ambiguous for AI agents to understand them. Jswarm does not offer that richness and does not suffer from that ambiguity, because Jswarm is in a form close to first-order predicate logic and so is easily interpreted by AI agents. Nevertheless, Jswarm is close enough to natural language, for humans to understand it.
• JSwarm has free word order. This offers robust communication with unchanged meaning even when the order changes. The paper mentions free word order in Jingulu, but I didn’t see an example of this feature in Jswarm or any explanation of how JSwarm uses free word order.  

Conclusion

I don’t understand anywhere enough about swarms of AI agents to identify the pros and cons of JSwarm or to evaluate the claims made for it. But certainly, this human understood the specimens of JSwarm sentences without difficulty, so at least that part of the claim seems to stand up to gentle scrutiny.  

The paper continues some interesting material on Jingulu, a language I hadn’t come across before. I think, though, that the press have slightly misunderstood in claiming that Jingulu has only 3 verbs. The language seems to have plenty of verbal expressions, it just splits them syntactically into a component expressed with a verb (which expresses tense and direction) and a component expressed with a root (which does not express tense and direction).