Date: Thursday 16th March
Room: College Building C133
Dr Joe Levy (University of Roehampton)
'Why Statistical Representations of Word Meaning are Interesting for Psychology'
For the last twenty years there has been increasing interest in the use of computational methods to capture certain aspects of word meaning. These methods capture how words are used by looking at the other words that they co-occur with, building patterns of co-occurrence or “semantic vectors” that can be used to measure the semantic distances between words. The techniques used have been driven by technological concerns but there has always been an interest in them from some psychologists. I will describe some current techniques and argue that semantic vectors may be able to play an important role in computational/statistical models of human behaviour. I will illustrate these points by examining the use of semantic vector methods in solving vocabulary multiple-choice tests.
I am a cognitive scientist with interests in language, memory and social cognition. I have used techniques from computational modelling, cognitive psychology and cognitive neuroscience.
Along with colleagues from Roehampton, recent projects have included:
the use of experimental methods, EEG and fMRI to examine action observation and perspective taking; the use of fMRI to measure the association of the default mode network and measures of empathy; accounting for variation in children’s reading performance with measures of metacognition.
A current focus is work with my colleague John Bullinaria (University of Birmingham) that continues our long collaboration of working on computational measures of word meaning. Our techniques have been very successful in generating numerical representations of the patterns of usage of different words in large bodies of text. The differences or distances between these "semantic vectors” can be shown to reflect the semantic relationships between different words. Recently, we have applied our technique to successfully improve models of cortical activation during word meaning processing tasks. Currently, along with Dr Samantha McCormick at Roehampton, we are looking at the various ways our semantic vectors can explain the linguistic structure of and human performance on vocabulary multiple choice tests. Plans for the future include exploring further applications of these semantic representations in the modelling of linguistic, cognitive and neuroscientific phenomena.