Our amazing Data Scientist Henrietta Eyre recently had a research paper published in Adaptive Behavior, the official journal of the International Society of Adaptive Behavior. Partnering with Jonathan Lawry of the University of Bristol, Henrietta explores language games with vague categories and negations. Below is the abstract and introduction, with a link at the end to access the full article.


The guessing game is a language game which models the communication protocols between two agents who aim to match category labels with objects encountered in a simulated or real environment.  Here we present a new representation framework for guessing games, in which category definitions explicitly incorporate semantic uncertainty and typicality. More specifically, we propose a conceptual model based on prototype and random set theory, in which categories are defined within a metric conceptual space. We argue that this conceptual framework is both expressive and naturally generates robust assertion and concept updating models. In particular, we define both assertion and updating rules for guessing games involving a mixture of labels and negated labels. Finally, the results of language game simulations are presented, where a multi-agent system evolves through pairwise language games incorporating an assertion and an updating algorithm.

Our results suggest that, within this framework, a mixture of both positive and negative assertions may be required in order for agent interpretations to converge, whilst retaining sufficiently discriminatory categories for effective communication.


Language games offer a model under which artificial systems may take account of the evolutionary nature of language learning. In a population of agents playing language games, the syntactic and semantic structures may be individual to each agent, but groups of agents then cooperate to evolve common structures and methods of communication. The underlying hypothesis then is that an evolutionary approach to determining a shared language will result in communications which are more efficient than if a designer attempted to preprogram agents with a fixed syntax and semantics. In this paper we present an investigation of the guessing game.

We will investigate the emergence of categories which are inherently vague. Here we associate vagueness with blurred boundaries and adopt a probabilistic approach consistent with those proposed by Hisdal, Edgington, Lawry, Bennett, Lassiter, and Goodman and Lassiter. From this perspective vagueness is seen as the result of uncertainty about the correct definition of categories which naturally arises as an integral part of the concept learning process. Spranger and Pauw have presented extensive investigations into vague categories in language games. Our conceptual model also describes vague categories with uncertain category boundaries, but also allows for categories to be situated in a conceptual space so that there may be a region of the conceptual space whose conceptual definition is not covered by any category.

Under this conceptual model we present a new investigation to the literature, where agents may describe the focus of a language game using negated labels as well as basic label descriptions. We argue that it is both realistic and useful to adopt a this richer assertion set.

You can read and download the full article here.

An outline of the paper is as follows:

Section 2 – Language Games. This section gives an overview of the language games literature.

Section 3 – Modelling Concepts. This section discusses cognitive models, conceptual space model and the random set and prototype model of concepts which we use as the representation framework for our language games.

Section 4 – Concept Evolution in a Language Game. This section gives a description of our language game model, including assertion and updating algorithms.

Section 5 – Measures of Convergence and Overlap. This section describes the metrics which quantify the behaviour of the language game system during the experimental simulations.

Section 6 – Experiments. This section presents the results from our simulation studies.

Section 7 – Conclusions. This final section gives some discussion and conclusions.

Ettie is a trained mathematician with a PhD in artificial intelligence. After a stint in academia she found her way to industry and now works on crunching natural language processing algorithms at Black Swan.