Linguistics and the changes that occur to languages over time are difficult to explain, but artificial intelligence studies could open doors to new theories on the development of natural languages.
According to this article, the scientists of the artificial intelligence divisions within Facebook and Google, together with the University of New York, have developed a framework in which artificial intelligence agents found themselves having to solve puzzles and of games and in the process have applied some linguistic phenomena typical of natural language.
The AI agents in question were trained with reinforced deep learning techniques, i.e. through a rewards system designed to guide them towards certain goals. It is not the first time that especially Facebook investigates the linguistic phenomena using machine learning algorithms: in 2017, in fact, had studied a system formed by two AI agents that communicated with each other through chat messages.
The new system studied with the University of New York, however, is more advanced in terms of the use of new generation agents capable of processing very complex neural networks capable of handling “rich perceptive inputs”. In this way the researchers showed how IA agents were able to learn from the language of their interlocutors and to improve their ability to express themselves.
The research team introduced groups of agents within a simulated environment. Agents were characterized by different types of skills: from solving simple equations to managing complex neural networks. To the agents were then given data of the games with specific characteristics: they were symmetrical, to allow the agents to act both as “speakers” and “listeners”; and required interaction with external elements that are part of the scenario.
After numerous experiments, the success rate between attempts to deal with the game on their own and those in which agents dealt with puzzles as a team was comparable. In a scenario of this type, a common language emerges that is shared among all the agents in the case in which there is a minimum number of agents that from the beginning master that language.
In a subsequent test, two different sets of linguistic agents were put in a position to interact with each other. The researcher’s report that all the agents have learned to “talk” with the agents of the other community and that, moreover, some agents, called “bridge” agents, have managed to manage the new shared protocol more effectively than the others.
According to the outcome of the study, in cases where there was a high communication rate within the same group, it was easier to solve the game. If the interaction within the group, instead, occurred less frequently than the interaction between different groups then the success rate of the game decreased. “This discovery demonstrates the rapid transition to a shared protocol between both groups when a reference language emerges within the group, regardless of whether they actually interact with the agents of the other group”.
In other words, a contact language is created between the groups when the agreement for a shared language is found within the group. In all scenarios, agents preferred to integrate or assimilate, rather than segregate, when speaking of language. Moreover, linguistic complexity tended to be greater proportionally to the difference in the number of agents between the two groups. “This means that two new languages resulting from the interaction of two communities of similar size tend to be substantially simpler,” the researchers concluded.
These results suggest that language does not depend on advanced and complex linguistic skills, say the authors of the research, but it can arise from “simple social exchanges” between “agents with perceptive capacities” who are encouraged to set up communication games. “We have observed that a symmetric communication protocol emerges without any innate and explicit mechanism built into an agent. The only requirement is that there are at least three agents of any kind of linguistic community,” they wrote.
“With a sufficient level of communication within the group, the languages become intelligible even among different groups, even for agents who have not been directly exposed to the other language”, the researchers conclude. “Over time a dominant language is formed that leads to the extinction of the other language. Alternatively, we are witnessing the origin of an original” Creole “protocol that has a lower complexity than the original languages, in case the two groups are populated by a similar number of agents”.Tags: AI, artificial intelligence, Facebook, Google