Fig 1.
The three-layer neural network architecture of a skip-gram model. An example input word, families, is depicted with a one-hot representation as well as example weight coefficient matrices used to produce an output value prediction for the word happy.
Fig 2.
Histogram of course equivalency scores.
Distribution of course equivalency validation scores from 400 model experiments, calculated as the median similarity rank of each pair. Lower is better. Scores > 50 omitted due to high skew.
Table 1.
Best models by overall validation score (median rank) for each validation set and cross-list collapse percentage.
Fig 3.
PCA of vector offsets with a Sequence and Honors constellation depicted using Physics courses.
Table 2.
A selection of analogy results from each of the six relationship categories.
Fig 4.
Conceptual decompositions of courses in the Subjects of (A) Mathematics and Education, (B) Economics, Public Policy, and Statistics and (C) all subject vectors.
Table 3.
Exemplar subject compositions.
Fig 5.
(A) all course vectors with close-ups of the departments of (B) History and (C) Near Eastern Studies.
Fig 6.
Close-up of race & gender studies cluster.
Fig 7.
Close-up of Asian languages & culture cluster.
Table 4.
Results of predicting attributes from course vectors.
Table 5.
Rules for preprocessing the semantic model training corpus.
Table 6.
Semantic model descriptions of subject vectors using biases of 0.5 and 1.
Table 7.
Semantic model description of missing subjects, the origin vector, and course vector differences (0.5 bias).