Uncovering the organization of semantic structure with similarity and inductions
Abstract/Contents
- Abstract
- Can semantic representations be captured by explicit structures? While connectionist models have long presented a challenge to rigidly structured models, recent work using structured probabilistic models have reintroduced such representations. For example, it has been proposed that trees are an appropriate choice of structure for capturing semantic judgments about blank biological properties of animals. In this work, we call this claim into question. We examine three models that have been explored in recent work on semantic cognition, including a structured probabilistic model, a neural network model, and a null model based on the raw covariance matrix. The models are compared to property induction judgments and similarity ratings across several sets of studies. First, we replicate a previous success of the structured model on a two-premise induction over mammals, but show that the connectionist model can also succeed at modeling this task. Second, we demonstrate that the structured model rapidly loses effectiveness when tested on sets of mammals which include cross-tree similarities such as living in water. Third, we examine the influence of question context on induction performance, by asking about diet, habitat, and behavior as well as blank biological properties. Finally, we extend beyond mammals to include birds and fish, gathering a rich set of similarity judgments, and demonstrate that these similarity judgments include systematic structure which no tree could capture. These results confirm that tree structures can serve as a good first-order representation of animal similarities and property induction judgments. However, they also show that these highly structured representations fail to capture subtle, yet important, effects in semantic judgments, specifically those that are not consistent with the tree structure. Other models which do not impose specific inductive biases on their training data, such as connectionist models or the null model, can capture these subtler effects as well.
Description
Type of resource | text |
---|---|
Form | electronic; electronic resource; remote |
Extent | 1 online resource. |
Publication date | 2011 |
Issuance | monographic |
Language | English |
Creators/Contributors
Associated with | Glick, Jeremy Jacobo |
---|---|
Associated with | Stanford University, Department of Psychology |
Primary advisor | McClelland, James L |
Thesis advisor | McClelland, James L |
Thesis advisor | Boroditsky, Lera |
Thesis advisor | Thomas, Ewart A. C |
Advisor | Boroditsky, Lera |
Advisor | Thomas, Ewart A. C |
Subjects
Genre | Theses |
---|
Bibliographic information
Statement of responsibility | Jeremy Jacobo Polanshek Glick. |
---|---|
Note | Submitted to the Department of Psychology. |
Thesis | Thesis (Ph.D.)--Stanford University, 2011. |
Location | electronic resource |
Access conditions
- Copyright
- © 2011 by Jeremy Jacobo Glick
- License
- This work is licensed under a Creative Commons Attribution Non Commercial 3.0 Unported license (CC BY-NC).
Also listed in
Loading usage metrics...