Language Evolution and Computation Bibliography

For robots to effectively bootstrap the acquisition of language, they must handle referential uncertainty-the problem of deciding what meaning to ascribe to a given word. Typically when socially grounding terms for space and time, the underlying sensor or representation was specified within the grammar of a conversation, which constrained language learning to words for innate features. In this paper, we demonstrate that cross-situational learning resolves the issues of referential uncertainty for bootstrapping a language for episodic space and time; therefore removing the need to specify the underlying sensors or representations a priori. The requirements for robots to be able to link words to their designated meanings are presented and analyzed within the Lingodroids-language learning robots-framework. We present a study that compares predetermined associations given a priori against unconstrained learning using cross-situational learning. This study investigates the long-term coherence, immediate usability and learning time for each condition. Results demonstrate that for unconstrained learning, the long-term coherence is unaffected, though at the cost of increased learning time and hence decreased immediate usability.

Time and space are fundamental to human language and embodied cognition. In our early work we investigated how Lingodroids, robots with the ability to build their own maps, could evolve their own geopersonal spatial language. In subsequent studies we extended the framework developed for learning spatial concepts and words to learning temporal intervals. This paper considers a new aspect of time, the naming of concepts like morning, afternoon, dawn, and dusk, which are events that are part of day-night cycles, but are not defined by specific time points on a clock. Grounding of such terms refers to events and features of the diurnal cycle, such as light levels. We studied event-based time in which robots experienced day-night cycles that varied with the seasons throughout a year. Then we used meet-at tasks to demonstrate that the words learned were grounded, where the times to meet were morning and afternoon, rather than specific clock times. The studies show how words and concepts for a novel aspect of cyclic time can be grounded through experience with events rather than by times as measured by clocks or calendars.

We present a series of studies investigating the formation, generative power, and evolution of toponyms (i.e. topographic names). The domain chosen for this project is the spatial concepts related to movement through the environment, one of the key sets of concepts to be grounded in autonomous agents. Concepts for spatial locations cannot be directly perceived and require representations built from interactions and inferred from ambiguous sensory data. A generative toponymic language game has been developed to allow the agents to interact, forming concepts for locations and spatial relations. The studies have shown that a grounded generative toponymic language may form and evolve in a population of agents interacting through language games. Initially, terms are grounded in simple spatial concepts directly experienced by the robots. The generative process then enables the robots to learn about and refer to locations beyond their direct experience, enabling concepts and toponyms to co-evolve.

A set of simulations are presented that investigate generalization in languages evolved for mobile robots. The mobile robot platform is RatSLAM, a model for Simultaneous Localization and Mapping based on rodent hippocampus that uses visual and odometric information to build up a map of the explored environment. The language agents use information from this system as inputs and are based on simple recurrent neural networks. This paper describes two sets of experiments exploring the nature of generalization in evolved languages. The first study investigated languages evolved from visual inputs and the second study investigated languages evolved from position representations. These studies showed that processing the input prior to the language agent affects the expressivity of the languages and the performance of the agents. Some generalization occurs in these languages. Studies are ongoing to extend these simulations using the simulated world of the robots.

We present a framework and first set of simulations for evolving a language for communicating about space. The framework comprises two components: (1) An established mobile robot platform, RatSLAM, which has a 'brain' architecture based on rodent hippocampus with the ability to integrate visual and odometric cues to create internal maps of its environment. (2) A language learning system based on a neural network architecture that has been designed and implemented with the ability to evolve generalizable languages which can be learned by naive learners. A study using visual scenes and internal maps streamed from the simulated world of the robots to evolve languages is presented. This study investigated the structure of the evolved languages showing that with these inputs, expressive languages can effectively categorize the world. Ongoing studies are extending these investigations to evolve languages that use the full power of the robots representations in populations of agents.

Using computational modeling techniques, this paper explores the range of conditions under which structured, language-like communication systems can emerge. In particular, we reconsider Simon Kirby's learning bottleneck model of linguistic adaptation using a different learning mechanism and different semantic domain. We demonstrate how parameters such as population size and training corpus size affect the likelihood of a population reaching consensus on a structure communication system.

Traditional theories of child language acquisition center around the existence of a language acquisition device which is specifically tuned for learning a particular class of languages. More recent proposals suggest that language acquisition is assisted by the evolution of languages towards forms that are easily learnable. In this paper, we evolve combinatorial languages which can be learned by a simple recurrent network quickly and from relatively few examples. Additionally, we evolve languages for generalization in different ``worlds'', and for generalization from specific examples. We find that languages can be evolved to facilitate different forms of impressive generalization for a minimally biased learner. The results provide empirical support for the theory that the language itself, as well as the language environment of a learner, plays a substantial role in learning: that there is far more to language acquisition than the language acquisition device.

We develop a new framework for studying the biases that recurrent neural networks bring to language processing tasks. A semantic concept represented by a point in Euclidian space is translated into a symbol sequence by an encoder network. This sequence is then fed to a decoder network which attempts to translate it back to the original concept. We show how a pair of recurrent networks acting as encoder and decoder can develop their own symbolic language that is serially transmitted between them either forwards or backwards. The encoder and decoder bring different constraints to the task, and these early results indicate that the conflicting nature of these constraints may be reflected in the language that ultimately emerges, providing important clues to the structure of human languages.