Computational accounts of language evolution grapple with several central dilemmas, including:
1. Adaptivity versus communicability: How can a distributed language system adapt while maintaining communicability?
2. Material versus representation: How does the material substrate of language (sound, text, bits, molecules) come to serve as both object and representation?
3. Lexical versus structural semantics: What structures---lexicon, context, grammatical rules---(should) carry what meanings, and how?
4. Learnability versus functionality: How can a distributed linguistic system balance languages' ease-of-learning against their functional information carrying capacity?
5. Individuality versus collectivity: How do collective properties (e.g., sublanguage distributions, interaction toplogies) shape individuals; how do individual properties spread, fixate, or disappear?
This talk illustrates UIUC work on several of these. In the end, we find that language evolution models generalize nicely to several interesting domains.
Kiran Lakkaraju
Agents' convergence to shared ontologies, lexicons, and grammars are specific instances of the general Multi-agent Agreement Problem (MAP): reaching consensus with decentralized knowledge and action. MAPs can vary in several dimensions including the number and accessibility of points in the possible agreement space; possible interactions among agents; information available; fitness (intrinsic value) of agreement points, etc. Solution concepts, too, vary by interaction topology, how agents gather information, and how they apply it (e.g., update or learning rules). By casting language evolution problems as MAPs we situate them among a rich and diverse set of work that includes statistical physics, game theory, norms/conventions, flocking/swarming, and control theory. Even better, language convergence stretches existing MAP theory because language requires multi-level agreement: agents must simultaneously agree upon different but interacting language aspects with different configurations (e.g. lexicon and grammar).
We show how to map linguistic convergence problems into a MAP framework, and then highlight how the algorithms developed for these varying fields can influence problems in language evolution, and vice versa.
Simple languages spread easily through a population by virtue of being easily learnable. Similarly, complex languages spread through a population by virtue of being more functional. In a population of learning agents, by treating their hypothesis class as a channel, we can connect the agents' language to the task they are trying to solve. We can then show that population dynamics result in learned languages and problem solutions that are simple but not too simple. In effect, population dynamics provide an implicit Occam's razor.
"Semiotic dynamics" is a very general issue; studying it for human (or machine agents') information processes is just the tip of a very interesting iceberg. Among other arenas, it is specifically an essential issue for biology:
"A common thread that links language and multicellularity is communication (interaction at a distance). In each case a complex, sophisticated network of interactions forms the medium within which the new level of organization (entities) comes into existence. The advent of translation can be seen similarly. Translationally produced proteins, multicellular organisms, and social structures are each the result of, emerge from, fields of interaction when the latter attain a certain degree of complexity and specificity. In the first case, we speak of tRNA "adaptors"; in the second, of morphogenetic fields (17); in the third, of language. Communication, networking, and discrimination are all buried deep in the evolutionary dynamic." [2]
"The biological world depends in its every process on the transmission of information...from mother cell to daughter cell, from one cell or tissue type to another, from one generation to the next, and from one species to another....[this] bioinformatics is the fundamental core of biology." [1]
This presentation describes several preliminary results from our work on applying insights from language evolution models to collective dynamics of bio-representation and bio-communication in molecular, genetic, intra/inter-cellular, and immune systems.
References
[1] David Eisenberg, Edward Marcotte, Andrew D. McLachlan, and Matteo Pellegrini, "Bioinformatic challenges for the next decade(s)," Philosophical Transactions of the Royal Society B, 2006.
[2] Woese, Carl R. "A new biology for a new century." Microbiology and Molecular Biology Reviews. 2004; 68:173--186.