A Tale of Two Semantic Systems: Taxonomic and Thematic Knowledge

A Tale of Two Semantic Systems: Taxonomic and Thematic Knowledge Daniel Mirman (mirmand@einstein.edu) Grant M. Walker (walkerg1@einstein.edu) Kristen M. Graziano (graziank@einstein.edu) Moss Rehabilitation Research Institute 50 Township Line Rd., Elkins Park, PA 19027, USA Abstract Behavioral, neuroimaging, and lesion analysis data suggest two parallel semantic systems. One system, with anterior temporal lobe as critical hub, captures taxonomic relations based on feature overlap. A second system, with temporo- parietal junction as critical hub, captures thematic relations based on complementary roles in events. We describe a computational model of this theory that accounted for a one- way behavioral dissociation in aphasic picture naming errors (more taxonomic errors than thematic errors) and a neuroanatomical double dissociation (damaging feature representations led to relatively more taxonomic errors, damaging event representations led to relatively more thematic errors). The model also predicted that both taxonomic and thematic competitors should be automatically activated during single word processing, with taxonomic competitors activated more quickly and more strongly. These predictions were tested and confirmed in a spoken word comprehension experiment using eye tracking to assess the time course of competitor activation. Keywords: semantic knowledge; taxonomic relations; thematic relations; event processing; computational modeling; spoken word processing. Introduction A core question in cognitive science is how semantic knowledge is represented. The study of semantic knowledge is typically intertwined with the study of feature-based or hierarchical conceptual categories. Feature-based accounts can explain a very broad range of phenomena (e.g., Rogers & McClelland, 2004) and are particularly effective in capturing the categorical, or taxonomic, structure of conceptual knowledge (e.g., Rogers & McClelland, 2004; O’Connor, Cree, & McRae, 2009). However, thematic conceptual knowledge – the grouping of concepts by participation in the same scenario or event (e.g., Estes, Galonka, & Jones, 2011) – is not as well captured by traditional feature-based accounts. On feature-based accounts, semantic similarity is a function of feature overlap (e.g., Cree, McRae, & McNorgan, 1999; Mirman & Magnuson, 2009; Rogers & McClelland, 2004), but thematically related objects typically share few, if any, features. Rather, they have complementary features that are related to the complementary roles the objects play in events. There is a long history of behavioral studies demonstrating that thematic knowledge plays an important role in adult conceptual knowledge (e.g., Goldwater, Markman, & Stilwell, 2011; Hare et al., 2009; Lin & Murphy, 2001; for a review, see Estes et al., 2011). One recent study used voxel-based lesion-symptom mapping (VLSM) to examine the neural basis of taxonomic and thematic processing (Schwartz et al., in press). Schwartz et al. analyzed picture-naming errors generated by a large sample of individuals with left hemisphere stroke aphasia (N=86), distinguishing between taxonomic errors (e.g., apple named as “pear” or “grape”) and thematic errors (e.g., apple named as “worm” or dog named as “bone”). Taxonomic errors were defined as category coordinates, superordinates, or subordinates; thematic errors were defined as incorrect responses which come from a different category but frequently play a complementary role with the target in events. The behavioral results showed a single dissociation: there were far more taxonomic errors than thematic errors (approximately 5:1 ratio) and all but two patients made more taxonomic errors than thematic errors. However, the lesion analysis revealed a neuroanatomical double dissociation in the relative proportion of taxonomic to thematic errors. Patients with lesions affecting the left anterior temporal lobe (ATL; Brodmann area 38 and the anterior portions of BA 20 and 21) tended to produce a higher proportion of taxonomic errors relative to thematic errors. In contrast, patients with lesions affecting the left temporo-parietal junction (TPJ; BA 39, posterior BA 21 and 22, superior BA 37, and BA 41 and 42) tended to produce a higher proportion of thematic errors relative to taxonomic errors. The ATL effect is consistent with previous studies demonstrating its critical role in lexical-semantic processing (e.g., Hodges, Graham, & Patterson, 1995; Lambon Ralph et al., 2001; Patterson, Nestor, & Rogers, 2007; Schwartz et al., 2009). The TPJ effect makes contact with studies suggesting an important role for TPJ in thematic relations (e.g., Kalenine et al., 2009) and relational knowledge more generally (e.g., Wu, Waller, & Chatterjee, 2007; for a recent comprehensive review of neuroimaging studies of semantic representations see Binder et al., 2009). Our first goal was to develop a formal computational model of these complementary semantic systems that can account for the neuroanatomical double dissociation as well as the one-way behavioral dissociation. Our model is related to previous work by Plaut (1995), who distinguished between semantic relatedness based on semantic feature overlap and semantic association based on temporal co-