What is semantic diversity and why does it facilitate visual word recognition?

Abstract

Previous research has speculated that semantic diversity and lexical ambiguity may be closely related constructs. Our research sought to test this claim in respect of the semantic diversity measure proposed by Hoffman et al. (2013). To this end, we replicated the procedure described by Hoffman et al., Behavior Research Methods, 45(3), 718-730 (2013) for computing multidimensional representations of contextual information using Latent Semantic Analysis, and from these we derived semantic diversity values for 28,555 words. We then replicated the facilitatory effect of semantic diversity on word recognition using existing data resources and observed this effect to be greater for low-frequency words. Yet, we found no relationship between this measure and lexical ambiguity effects in word recognition. Further analysis of the LSA-based contextual representations used to compute Hoffman et al. (2013) measure of semantic diversity revealed that they do not capture the distinct meanings of ambiguous words. Instead, these contextual representations appear to capture general information about the topics and types of written material in which words occur. These analyses suggest that the semantic diversity metric previously proposed by Hoffman et al. (2013) facilitates word recognition because high-diversity words are likely to have been encountered no matter what one has read, whereas many participants may not have encountered lower-diversity words simply because the topics and types of written material in which they occur are more restricted.

Keywords: Latent semantic analysis; Lexical ambiguity; Semantic diversity; Word frequency.

Fig. 1

Illustration of semantic diversity procedure

Fig. 2

Scatter plots of resulting semantic diversity measures on x-axes and the norms reported by Hoffman et al. (2013) and the y-axes. On the left, values obtained following the preprocessing procedure described in the methods (lemmatised corpus, exclusion of stop words, etc.), while on the right, values obtained following Hoffman et al. (2013) preprocessing procedure. On the top row (blue) are the measures obtained with the classical output of LSA (weighting by the singular values), while on the bottom row (grey) are the measures obtained without considering the singular values

Fig. 3

Model estimates of the effect of semantic diversity by frequency on reaction time data as a function of database

Fig. 4

Results of the simulation analysis of Experiment 1 of Rodd et al. (2002) on reaction time data of BLP and ELP. Both datasets show that increasing number of senses speeds performance, while increasing number of meanings slows performance

Fig. 5

Results of the simulation analysis of Experiment 1 of Rodd et al. (2002) showing no difference in semantic diversity for words with many or few senses and meanings

Fig. 6

Results of the simulation analysis of Experiment 2 of Rodd et al. (2002) on response time data from the BLP and ELP. Data show that an increased number of senses speeds lexical decision latency, but that there is no effect of the number of meanings

Fig. 7

Results of the simulation analysis of Experiment 2 of Rodd et al. (2002) showing no difference in semantic diversity values for words with many or few senses or meanings

Fig. 8

Descriptive bar plots of response time data (left) by type of ambiguity (pooled between all experimental conditions) as reported by Armstrong & Plaut (2016) and bar plots of replication analysis of BLP and ELP (middle and right, respectively) showing a polysemy advantage but no homonymy disadvantage

Fig. 9

Results of the simulation analysis of Armstrong & Plaut (2016) on semantic diversity measures showing no difference across ambiguity type

Fig. 10

t-SNE plots of the context vectors in which the word calf occurs

Fig. 11

t-SNE plots of the context vectors in which the word mole occurs

Fig. 12

t-SNE plots of the context vectors in which the word pupil occurs

Fig. 13

t-SNE plots of the whole corpus labelled by domain on the top (variance ratio: 320.74), while on the bottom are the same labels randomly assigned for comparison (M = 1.00, SD = 0.04 for 1000 iterations)

Fig. 14

t-SNE plots of the whole corpus labelled by type of written material on the top (variance ratio: 301.12), while on the bottom are the same labels randomly assigned for comparison (M = 1.00, SD = 0.05 for 1000 iterations)