Disentangling narrow and coarse semantic networks in the brain: The role of computational models of word meaning

Abstract

There has been a recent boom in research relating semantic space computational models to fMRI data, in an effort to better understand how the brain represents semantic information. In the first study reported here, we expanded on a previous study to examine how different semantic space models and modeling parameters affect the abilities of these computational models to predict brain activation in a datadriven set of 500 selected voxels. The findings suggest that these computational models may contain distinct types of semantic information that relate to different brain areas in different ways. On the basis of these findings, in a second study we conducted an additional exploratory analysis of theoretically motivated brain regions in the language network. We demonstrated that data-driven computational models can be successfully integrated into theoretical frameworks to inform and test theories of semantic representation and processing. The findings from our work are discussed in light of future directions for neuroimaging and computational research.Keywords LSA . HAL . Semantic space models . Coarse semantic coding . fMRI Latent semantic analysis (LSA; Landauer & Dumais, 1997) and the hyperspace analogue to language (HAL; Lund & Burgess, 1996) are among the most influential computational models of word meaning. LSA and HAL, among other socalled Bsemantic space models^or Bdistributional semantic models,^use word co-occurrence frequencies as the basic building blocks for word meaning (see Jones, Willits, & Dennis, 2015, for a recent review). In these models, the cooccurrence frequencies of a word with all the other documents (as in LSA) or all other words with which the word occurs (as in HAL) are used to build the vector representation for that word, typically based on a very large-scale text corpus. The resulting representation of any target word is a highdimensional vector with each dimension denoting either a word (word-to-word matrix) or a document (word-to-document matrix). The raw vectors may consist of thousands or tens of thousands of dimensions and are usually very sparse. Dimension reduction methods are often used to reduce the number of dimensions in these models. These standard methods used by LSA and HAL have since been further developed or expanded. For example, probabilistic LSA (Hoffman, 2001) and its fully Bayesian extension the Topic model (Griffiths, Steyvers, & Tenenbaum, 2007) can identify lexemes with multiple senses (Tomar et al., 2013) and generate semantic representations as probability distributions rather than points in a high-dimension space. Positive pointwise mutual information (PPMI) has been used in place of raw cooccurrence frequencies (Bullinaria & Levy, 2007). Zhao, Li, and Kohonen (2011) integrated these models into a selforganizing map framework, and Fyshe, Talukdar, Murphy, and Mitchell (2013) discussed how different types of constraints on what counts as a co-occurrence qualitatively affect semantic information. Evaluation of semantic space modelsComputational models...

Speaking two “Languages” in America: A semantic space analysis of how presidential candidates and their supporters represent abstract political concepts differently

“…For the USF norms and the MTurk data, the number indicates the word association strength, calculated as the proportion of participants who produced the target word in response to a cueparticipants were asked to write down a word that came to mind (e.g., Bline^) when presented with a prompt word (e.g., Bborder^) ** USF = the word associations based on the University of South Florida free word association norms of (Nelson et al, 1998). The USF norms include only single words and did not include Bhealth care* ** MTurk = the word associations based on 600 Mechanical Turk responders of (Schloss et al, 2016). The MTurk data include only single words and did not include Bhealth care1 Study Two: Voters' self-reported political engagement and their voting behavior Study One was primarily concerned with how the presidential candidates and their parties represented political concepts and the changing representations over time.…”

“…In our modeling, we trained the model using both algorithms as implemented in Python's Gensim package (Rehurek, 2010) and concatenated the representations into vectors with 4,000 dimensions. The decision to use concatenated vector representations was based on the consideration that (a) the two algorithms may be sensitive to different types of word associations (e.g., dominant, paradigmatic, vs. non-dominant associations), which may implicate different processing mechanisms (Jung-Beeman, 2005), and (b) previous work has suggested that concatenated vectors can in some cases provide increased accuracy in representing subtle semantic differences (Fyshe et al, 2013;Schloss & Li, 2016). 1 We used word2vec's default settings, with a window size of five and a minimum word-count of five (words that were used less than five times were excluded) for all models reported below except for the time-course analysis.…”

“…The close correspondence between word meanings and language usage has motivated a new round of thinking about how computational models can capture individual differences in reading, writing, understanding, and knowledge in general (Landauer et al, 2013). Not surprisingly, computational semantic space models have also been successfully applied to account for patterns of brain activation, with a great deal of predictive accuracy and focal precision (Bullinaira & Levy, 2013;Mitchell et al, 2008;Schloss & Li, 2016), indicating that this data-driven approach is both psychologically and neurobiologically plausible. In a recent computational study, Caliskan et al (2017) used a word-embedding model (GloVe, very similar to word2vec used in this study; see Method under Study One) and showed that the model can capture implicit human biases: semantic space vectors generated from the GloVe model's text analyses showed stronger similarities (as measured by cosine similarity scores) between pairs of words that are stereotyped human associations, for example, between male names and words related to career versus female names and words related to family, and between male terms and science versus female terms and arts.…”

The “Small World of Words” English word association norms for over 12,000 cue words

“…Representational similarity as measured by voxel analysis is also becoming increasingly important in neuro-imaging approaches that try to uncover the structure of semantic memory. Across a range of studies, the fMRI evidence indicates that the pattern of activation across different areas of the brain when reading common words (Mitchell et al, 2008) can be predicted from distributional lexico-semantic models (Schloss & Li, 2016). Against this backdrop, it seems sensible to consider how the SWOW-EN norms might be used to measure semantic similarity.…”

Commonality of neural representations of sentences across languages: Predicting brain activation during Portuguese sentence comprehension using an English-based model of brain function

“…By contrast, semantic vector representations of concepts (e.g. Latent Semantic Analysis, (Dumais, 2004)) have limited linguistic interpretability, although the use of such representations has been shown to be predictive of neural activation (e.g., Mitchell et al, 2008;Murphy et al, 2012;Schloss & Li, 2016).…”