Mark Andrews’s research while affiliated with Nottingham Trent University and other places

Over the past 15 years, there have been two increasingly popular approaches to the study of meaning in cognitive science. One, based on theories of embodied cognition, treats meaning as a simulation of perceptual and motor states. An alternative approach treats meaning as a consequence of the statistical distribution of words across spoken and written language. On the surface, these appear to be opposing scientific paradigms. In this review, we aim to show how recent cross-disciplinary developments have done much to reconcile these two approaches. The foundation to these developments has been the recognition that intralinguistic distributional and sensory-motor data are interdependent. We describe recent work in philosophy, psychology, cognitive neuroscience, and computational modeling that are all based on or consistent with this conclusion. We conclude by considering some possible directions for future research that arise as a consequence of these developments.

We present an account of semantic representation that focuses on distinct types of information from which word meanings can be learned. In particu-lar, we argue that there are at least two major types of information from which we learn word meanings. The first is what we call experiential infor-mation. This is data derived both from our sensory-motor interactions with the outside world, as well as from our experience of own inner states, par-ticularly our emotions. The second type of information is language-based. In particular, it is derived from the general linguistic context in which words appear. The paper spells out this proposal, summarizes research supporting this view and presents new predictions emerging from this framework.

In Kousta, Vigliocco, Vinson, Andrews, and Del Campo (2011), we presented an embodied theory of semantic representation, which crucially included abstract concepts as internally embodied via affective states. Paivio (2013) took issue with our treatment of dual coding theory, our reliance on data from lexical decision, and our theoretical proposal. Here, we address these different issues and clarify how our findings offer a way to move forward in the investigation of how abstract concepts are represented.

Although much is known about the representation and processing of concrete concepts, knowledge of what abstract semantics might be is severely limited. In this article we first address the adequacy of the 2 dominant accounts (dual coding theory and the context availability model) put forward in order to explain representation and processing differences between concrete and abstract words. We find that neither proposal can account for experimental findings and that this is, at least partly, because abstract words are considered to be unrelated to experiential information in both of these accounts. We then address a particular type of experiential information, emotional content, and demonstrate that it plays a crucial role in the processing and representation of abstract concepts: Statistically, abstract words are more emotionally valenced than are concrete words, and this accounts for a residual latency advantage for abstract words, when variables such as imageability (a construct derived from dual coding theory) and rated context availability are held constant. We conclude with a discussion of our novel hypothesis for embodied abstract semantics.

In this paper, we describe a model that learns semantic representations from the distributional statistics of language. This model, however, goes beyond the common bag-of-words paradigm, and infers semantic representations by taking into account the inherent sequential nature of linguistic data. The model we describe, which we refer to as a Hidden Markov Topics model, is a natural extension of the current state of the art in Bayesian bag-of-words models, that is, the Topics model of Griffiths, Steyvers, and Tenenbaum (2007), preserving its strengths while extending its scope to incorporate more fine-grained linguistic information.

The authors identify 2 major types of statistical data from which semantic representations can be learned. These are denoted as experiential data and distributional data. Experiential data are derived by way of experience with the physical world and comprise the sensory-motor data obtained through sense receptors. Distributional data, by contrast, describe the statistical distribution of words across spoken and written language. The authors claim that experiential and distributional data represent distinct data types and that each is a nontrivial source of semantic information. Their theoretical proposal is that human semantic representations are derived from an optimal statistical combination of these 2 data types. Using a Bayesian probabilistic model, they demonstrate how word meanings can be learned by treating experiential and distributional data as a single joint distribution and learning the statistical structure that underlies it. The semantic representations that are learned in this manner are measurably more realistic—as verified by comparison to a set of human-based measures of semantic representation—than those available from either data type individually or from both sources independently. This is not a result of merely using quantitatively more data, but rather it is because experiential and distributional data are qualitatively distinct, yet intercorrelated, types of data. The semantic representations that are learned are based on statistical structures that exist both within and between the experiential and distributional data types.

In this paper we present models of how meaning is represented in the brain/mind, based upon the assumption that children develop meaning representations for words using two main sources of information: information derived from their concrete experience with objects and events in the world (which we refer to as attributional information) and information implic-itly derived from exposure to language (which we refer to as distributional information). In the first part of the paper we present a model developed using self-organising maps (SOMs) starting from speaker-generated features (properties that speakers considered to be important in defining and describing the meaning of a word). This model captures meaning similarity between words based solely upon attributional information and has been shown to be success-ful in predicting a number of behavioural semantic effects. In the second part of the paper, we present a probabilistic model that goes beyond attributional information alone, integrating this information with distributional information derived from text corpora. The ability of this integrated model to learn semantic relationships is demonstrated with reference to comparable probabilistic models that use only attributional or distributional information.

In recent studies of semantic representation, two distinct sources of information from which we can learn word meanings have been described. We refer to these as at-tributional and distributional information sources. At-tributional information describes the attributes or fea-tures associated with referents of words, and is acquired from our interactions with the world. Distributional in-formation describes the distribution of words across dif-ferent linguistic contexts, and is acquired through our use of language. While previous work has concentrated on the role of one source, to the exclusion of the other, in this paper we study the role of both sources in combi-nation. We describe a general framework based on prob-abilistic generative models for modelling both sources of information, and how they can be integrated to learn se-mantic representation. We provide examples comparing the learned structures of each of three models: attri-butional information alone, distributional information alone, and both sources combined.

In this paper, we generalize the models used by MacKay [1] in his analysis of evolutionary strategies that are based on sexual, rather than asexual, reproduction methods. This analysis can contribute to the understanding of the relative power of genetic algorithms over search methods based upon stochastic hill-climbing, e.g. [2], [3].

In this paper, we describe the problems of learning and inference in Hidden Stack Au-tomata. These are generalizations of Hidden Markov models (HMM), whereby the finite-state HMM now has access to a pushdown stack memory. Providing learning and infer-ence algorithms is greatly facilitated by de-scribing the system in terms of a dynamical Bayesian network that couples a HMM and a piecewise-linear dynamical system. This en-semble exactly implements a pushdown au-tomaton, with the dynamical system act-ing as the unbounded stack memory for the HMM. The contents of the stack are coded numerically as points in the state-space of the dynamical system, while operations on the stack contents are performed by way of its dynamics. We provide a sequential Monte-Carlo method for the inference of the state-space trajectories given a sequence of observations. This provides us with other-wise intractable posterior estimates of the stack-contents. From this we can derive an expectation-maximization (EM) algorithm to estimate the parameters of the system given our training data. We provide examples of learning and inference in these Hidden Stack Automata, and compare their performance to HMMs.