In their recent article, published open access in Psychological Reviewopens in new window, Hoffman and colleagues presented a computational model of the processes involved in acquisition and use of semantic memory.
Semantic memory refers to the vast database of conceptual knowledge we use to interpret language and the world around us. Contemporary theories hold that two types of information contribute to our object and lexical concepts: sensory-motor information derived from direct physical experience, and distributional data about how words are used together in language.
Understanding how these distinct sources of knowledge are integrated is a key challenge in cognitive science.
Hoffman et al's connectionist computational model provides insights into how this integration may be achieved. The model was trained with a set of 64 concepts and learned to map between words and their sensory-motor properties. An important concurrent training task for the model was prediction: words were presented in sequences and the model was trained to predict the next word in each sequence, thus extracting the underlying statistical regularities present in language.
Through these joint pressures, the model developed a single set of semantic representations that was simultaneously sensitive to similarities in physical properties and to similarities in word use in language.
A key novel finding was how the model learned to process abstract words. These words were never directly associated with any sensory-motor information, but the model did learn about their co-occurrences with other concrete words. Because of this, when the model was presented with abstract words, it came to activate the sensory-motor properties of the concrete objects with which it frequently co-occurred, despite never being explicitly trained to do so. So, for example, when it was presented with the abstract word "journey," it began to activate properties associated with vehicles.
These results suggest a potential mechanism by which the understanding of highly verbal, abstract concepts can become grounded in experience of the real world.
The second aim of the model was to advance the understanding of how concepts are retrieved and shaped to fit with current experience. This controlled processing of knowledge is critical because the same word can be used in very different ways in different situations.
The semantics of the word "bark," for example, changes depending on whether it is used in the presence of trees or dogs. To explore this element of semantic processing, Hoffman et al. presented their model with a multiple-choice task in which it was required to decide which concept was associated with a particular ambiguous word.
The model did well on trials that relied on understanding the most common interpretation of the word, especially when given a strong contextual cue, but it was less successful when the correct response was linked to a more unusual sense of the word.
To improve the model's performance, the researchers added a control process, which forced the network to be influenced by all of the possible choices while it processed the ambiguous word. This process guided the network toward an interpretation of the ambiguous word that was consistent with the choices on offer.
This mechanism may help us to understand how humans use cognitive control to shape the retrieval of concepts according to the constraints of their current environment.
Hoffman et al. tested the ability of the model to successfully complete three semantic tasks probing different types of conceptual relationships. Importantly, they also tested the performance of the model following damage, to simulate effects of neurological impairment.
When the model's conceptual representations were degraded, its performance closely simulated that seen in the syndrome of semantic dementia, in which a progressive loss of concepts is observed. A different performance profile emerged when the model's control processes were lesioned — here, performance closely matched that seen in stroke patients whose semantic deficits have been attributed to poor cognitive control.
These results suggest that the model provides an accurate account of how concepts are stored and used by both healthy and impaired populations.
Citation
- Hoffman, P., McClelland, J. L., & Lambon Ralph, M. A. (2018). Concepts, control, and context: A connectionist account of normal and disordered semantic cognition. Psychological Review, 125(3), 293–328. https://dx.doi.org/10.1037/rev0000094
Note: This article is in the Core of Psychology topic area. View more articles in the Core of Psychology topic area.