How groups develop a specialized domain vocabulary: A cognitive multi-agent model
Introduction
Languages evolve: like biological systems, they undergo mutation and selection as they are passed on between speakers and generations. Human communication as well as its biological analog evolve under environmental constraints. Fitness of a set of linguistic devices is, thus, also a function of the cognitive facilities. In this paper, we assume that the acquisition and retention of linguistic facts in memory is a crucial factor determining how languages are developed by communities. We use a cognitive architecture to provide an independently validated model of human memory to simulate the evolutionary process that produces a crucial part of the communication system: the vocabulary.
Recent models of dialogue describe how interlocutors develop representation systems in order to communicate; such systems can, for instance, be observed using referring expressions that identify locations in a maze. Experiments have shown that referring expressions converge on a common standard (Garrod & Doherty, 1994). Pickering and Garrod’s (2004) Interactive Alignment Model suggests that explicit negotiation and separate models of the interlocutor’s mental state are not necessary, as long as each speaker is coherent and adapts to his interlocutors, as speakers are known to do on even simple, linguistic levels (lexical, syntactic). This shifts the weight of the task from a sophisticated reasoning device to the simpler learning mechanism of the individual.
Some evolutionary models see the transmission of cultural information as a directed process, in which information is passed only from the older to the younger generation. Other models explain the emergence of language as a continuous process within generations. This process may be modeled as convergence towards the bias set by innate learning and processing systems of the individual, but it can also be seen as the result of ongoing changes that interact with the cultural environment of the collaborating language users. There, meaning–symbol connections spread between collaborating agents and ultimately converge on a predominant one. It is the dichotomy between individual and community-based learning that motivated the experiments by Fay, Garrod, Roberts, and Swoboda (2010) and Garrod, Fay, Lee, Oberlander, and Macleod (2007), which serve as the basis for the model presented here.
In the society of cognitive agents in Fay’s study and in our experiments, agents adapt their communication system collaboratively to the environmentally shaped and cognitively constrained needs of each individual. With our model, we aim to use a cognitive framework – specifically a memory model – to reflect processes in the individual that give rise to emergent convergence and learning within the community. By this, we acknowledge the fact that cultural evolution is constrained by individual learning; each agent learns according to their cognitive faculty (cf., Christiansen & Chater, 2008). The possibility of cultural language evolution between generations has been supported by computational simulations (e.g., Brighton et al., 2005, Kirby and Hurford, 2002). Kirby and Hurford’s (2002) Iterated Learning model of language evolution describes vertical development of a language system by feeding developed linguistic signs or conventions back into another agent. Thus, it abstracts away from the processes within the community that forms a generation, yet does not rely only on emergence through biological evolution of the system that processes language.
The individual language faculty as a result of biological evolution and adaptation to cultural language has been the focus of psycholinguistic models proposing specialized mechanisms (theChomskian viewpoint). While syntactic theory has long relied on production rule systems, more recent lexicalist approaches (Jackendoff, 1975) also integrate well with theories of general cognition (ACT-R: Anderson et al., 2004; SOAR: Laird & Rosenbloom, 1987). In this sense, the model presented here reflects the development of a common vocabulary, which we see as prototypical for that of the lexicon, the central component of a language.
Indeed, the multi-agent model discussed in the present paper sees part of the linguistic process as an instantiation of general cognition: the composition and retrieval of signs follows general cognitive mechanisms. Adaptation according to experience is determined by human learning behavior. Simulation in validated cognitive frameworks allows us to constrain the learning process by the bounds of human memory.
Griffiths and Kalish (2007), for instance, model language evolution through iteration among rational learners in a Bayesian framework; the purpose of the present project is to tie the simulation of language evolution to a concrete experiment and a more process-oriented cognitive architecture than the Bayesian framework. ACT-R’s learning mechanisms add a notion of recency (decay) to the Bayesian view. Work on language processing has modeled the relationship to ACT-R memory retrieval, both for language comprehension (Ball et al., 2007, Budiu and Anderson, 2002, Lewis et al., 2005, Stocco and Crescentini, 2005) and for language production (Reitter, 2008).
We introduce a cognitive model that simulates a participant in the experiment; multiple models interact as a community of participants. The purpose of this paper is to observe how a compositional vocabulary is created between collaborating agents in a computational cognitive simulation. Like Smith, Brighton, and Kirby (2003), we represent meaning–signal mappings using associations between memory items to create compositional signs, but we augment this model of pre-existing knowledge with one of explicitly encoded and retrievable domain knowledge. Other simulations have shown that cultural evolution leads to compositional languages (Kirby & Hurford, 2002).
We will show that the model demonstrates learning behavior similar to the empirical data. We assume these agents share a common reference system initially, display cooperative behavior and adopt mixed roles as communicators. Therefore, we explore different scenarios that test the necessity of our preconditions, in particular the fact that each agent can be both on the sending and the receiving end of the communications. The underlying question is whether dialogue (producing and comprehending language) is necessary for participants to establish joint communication. In search for factors that influence community convergence, we also examine the effect of initial common ground between agents and the role of the structure of the network that describes each agent’s knowledge. Specifically, we present results suggesting that the specific power-law distribution found in ontologies is beneficial to the within-community convergence.
Section snippets
The task
The Pictionary experiment (Garrod et al., 2007) involves two participants, a director, who is to draw a given meaning from a list of concepts known to both participants, and a matcher, who is to guess the meaning. Director and matcher do not communicate other than through the drawing shared via screens of networked computers; the matcher is able to draw as well, for instance to request clarification of a part of the picture. Each trial ends when the matcher decides to guess a concept. Garrod et
Architecture
ACT-R (Anderson, 2007) is an architecture for specifying cognitive models, one of whose major components is memory. ACT-R’s memory associates symbolic chunks of information (sets of feature–value pairs) with subsymbolic activation values. Learning occurs through the creation of chunks, which are then reinforced through repeated presentation, and forgotten through decay over time. The symbolic information stored in chunks is available for explicit reasoning, while the subsymbolic information
Simulation 1: learning and convergence
In the first simulation, we evaluate whether the model exhibits similar learning and convergence behavior, and whether there are differences in learning between the isolated–pair and community condition, as observed in Fay et al.’s experiment. The model uses the same number of concepts, trials and simulated participants as in the experiment. 100 repetitions of the simulation were run, each with a different, randomly sampled ontology structure; the same 100 ontologies were used for all
Simulation 2: director and matcher roles
Garrod et al. (2007) compared the performance of their participants in a comparable Pictionary task when a single director remained in that role throughout the experiment (single director, SD condition), vs. when participants swapped roles after each round (double director, DD condition). Identification accuracy was slightly higher for the role-switching, double-director condition than in the single-director condition (significantly so only in the final rounds 5 and 6). This experiment is
Simulation 3: noise in common ground
A vital assumption of the compositional semantics in this model is that the agents start out with some common knowledge. For instance, both director and matcher need to accept that ambulances and buildings are strongly related to the concept Hospital. However, the strength of each link between the same two concepts may differ between any two agents. This error does not necessarily preclude the matcher from making the right inference. The model allows us to test the role of inter-subject
Simulation 4: ontology structure
Does the structure of relationships between ontological concepts assist human communities in language convergence? In this task, human participants as well as cognitive models established new meanings of the concepts by combination. Accuracy of retrieval of target concepts given the combination of drawings depends on the ambiguity of the drawing–concept relations; in other words, it depends on how clearly the drawings identify the right target concept.
Commonly, the frequency of co-occurrences
Simulation 5: feedback
The game used in the experiments and simulations discussed here differs from the typical linguistic interaction between humans in one important aspect. Humans usually obtain some form of measure of success that determines whether the communication was received correctly. In the game, this measure of success is never explicit, and, at best, implied. For example, speakers may retroactively decide that a guess (of concept A) that they made earlier did not work, because a new sign much more clearly
General discussion
The model replicates several of the characteristics of the communities compared to the isolated pairs condition; specifically the setbacks after switching partners for the first few times and the ultimate convergence, despite very limited feedback. We also arrive at a clear prediction: bi-directionality is essential for linguistic convergence in communities. The model fails to explain several other characteristics of the data. While subjects gained most of their ID accuracy during round 1, the
Conclusion
We have demonstrated the use of validated, cognitively plausible constraints to explain an emergent, evolutionary group process via multi-agent simulation. Subsymbolic and symbolic learning within a validated human memory framework can account for rapid adaptation of communication between dyads and for the slower acquisition of a domain language in small speaker communities despite very limited feedback about the success of each interaction. Bi-directional communication is predicted to be
Acknowledgments
We thank Nicolas Fay and Simon Garrod for making their data and manuscripts available and Ion Juvina for comments. This work was funded by the Air Force Office of Scientific Research (FA 95500810356).
References (24)
- et al.
Language as an evolutionary system
Physics of Life Reviews
(2005) - et al.
Conversation, co-ordination and convention: An empirical investigation of how groups establish linguistic conventions
Cognition
(1994) - et al.
Soar: An architecture for general intelligence
Artificial Intelligence
(1987) - et al.
Syntactic comprehension in agrammatism: A computational model
Brain & Language
(2005) Rules of the mind
(1993)How can the human mind occur in the physical universe?
(2007)- et al.
An integrated theory of mind
Psychological Review
(2004) Word frequency distributions
(2001)- Ball, J., Heiberg, A., Silber, R. (2007). Toward a large-scale model of language comprehension in ACT-R 6. In R. L....
- et al.
Comprehending anaphoric metaphors
Memory & Cognition
(2002)