We investigate ways in which artificial agents can self-organize languages with natural-language like properties and how meaning can co-evolve with language. Our research is based on the hypothesis that language is a complex adaptive system that emerges through adaptive interactions between agents and continues to evolve in order to remain adapted to the needs and capabilities of the agents. We explore this hypothesis by implementing the full cycle of speaker and hearer as they play situated language games and observing the characteristics of the languages that emerge. For more information, visit the Evolutionary Linguistics website.
The language game framework is the basis of our work around the origins and evolution of language. The notion of language games was introduced by the philosopher Ludwig Wittgenstein to emphasize that the meaning of words, and of linguistic constructions in general, lies in how they are used in concrete activities or "games". In order to include this fundamental insight into our investigations, we implement language games played by robotic agents.
In a typical language game, two agents face a communication task. Each game is played within a certain context and with a certain goal. For instance, the context could be a number of color samples and the goal might be that one of the agents can identify a sample named by the other agent.
In a language game experiment, a large number of language games is played by different agents randomly selected from a population. All agents start out without a language, so that initially all games fail. Agents can learn from these failures, however, and become better at it if they manage to establish a shared language. With language game experiments we can thus investigate the learning and processing mechanisms needed for the establishment of a conventional language and get insight into the dynamics of conventionalization.
A good example of a language game is the Naming Game. A good introductory paper on the Naming Game is the following:
Steels, L., & Loetzsch, M. (2012). The Grounded Naming Game. In L. Steels (Ed.), Experiments in Cultural Language Evolution. Amsterdam: John Benjamins.
The semiotic cycle
During the course of a game, agents need to go through a number of phases, together making up aso called "semiotic cycle".
First, the sensory input needs to be conceptualized, so that meaningful parts are identified in it. Here, in line with Wittgensteinian principles again, "meaningful" means serving to accomplish the aim of the game. Thus, when the context consists of a number of color samples and the goal is to distinguish one of them, then a first step must be to "recognize" the samples as different things. This can be accomplished for instance by checking that all samples fall in different regions (equivalence classes) of color space (the space of possible RGB values), and by splitting up existing regions if not. In this case, the result of conceptualization will be a particular region in color space. In more extended setups, the resulting conceptualizations may have more structure, but the basic idea remains the same.
In the next phase, the speaker needs to verbalize his conceptualization. The simplest way to do this is by remembering "which word to use for it". In the example case, this would amount to associating color words to regions in color space, but more complicated (compositional) verbalizations are possible if conceptualizations have more structure. In contrast to conceptualizations, verbalizations or utterances can be transmitted between agents just as spoken, written or gestural language is public for humans.
After verbalization, the hearer agent has to interpret the speaker's utterance. This is a context dependent process and typically involves two sub phases. First, the utterance is parsed into a set of constraints on possible conceptualizations. These are then used to select among the conceptualizations supported by perception. The "winning" conceptualization then somehow must indicate the action that wins the game. In the example of color samples, the hearer may find that the color word is associated with a region color space that also contains one of the perceived color samples. The agent can then point to the sample identified in this way, and the speaker agent can confirm or correct, again through pointing.
Both agents then enter a final alignment or learning phase, in which they update their categories (the regions in color space) and language (the mapping from regions to color words) in order to optimize their success in future games.
Learning and Alignment
Many things can go wrong during a semiotic cycle. As already mentioned, an agent may not have the proper categories to make the relevant distinctions. In this case, agents extend their ontology by defining new categories. It may also happen that a speaker agent does not have the proper means to express a conceptualization. In this case, he may creatively search for alternatives to get the message through anyway. The hearer might then misunderstand the speaker, which can give rise to phenomena of grammaticalization etc.