Skip to main content icon/video/no-internet
icon backReturn to Entries

Semantic Memory, Computational Perspectives

Semantic memory refers to factual or conceptual knowledge that is not related to any given personal episode. For instance, the fact that Ottawa is the capital of Canada is a piece of knowledge that you may retain, without necessarily being able to identify when and where you learned it. Similarly, you may know that a poodle and a terrier are similar to each other without ever having been told that they are but rather by virtue of the fact that they fit together in a conceptual representation of the world because they are both small dogs.

The examples given above illustrate the two main kinds of semantic information—conceptual and propositional. A concept is a mental representation of something. So one might have the concept of a dog that becomes active when one sees a poodle, smells a Labrador, is licked by a terrier, or talks about greyhounds. Concepts can also include actions such as running and properties such as red or quickly. Propositions join concepts together into units of mental representation that are capable of having a truth-value. So the concept of a dog is neither true nor false. But the proposition that dogs have legs is typically true in our world, although one could imagine a world of legless dogs in which it was not true. Similarly, Ottawa is the capital of Canada, but it is easy to imagine a world in which Toronto was instead.

Early models of semantic memory relied on representations of conceptual and propositional knowledge that were supplied by the theorists. Starting in the late 1980s, however, attention turned to how knowledge could be extracted automatically from exposure to a corpus representing human experience. We will focus on these models starting with those that extract conceptual knowledge and then considering those that extract propositional knowledge.

Conceptual Knowledge

The earliest and most prominent of the models that extract conceptual knowledge is latent semantic analysis (LSA). Introduced by Scott Deerwester, Susan Dumais, George Furnas, Thomas Landauer, and Richard Harshman, LSA derives meaning using statistical computations applied to a large corpus of text. Semantically similar words tend to appear in similar documents. By observing which words appear in which documents in the corpus, LSA defines a set of mutual constraints. These constraints can be solved using singular value decomposition—producing vector representations of both words and documents. The similarity of these vectors is then used to predict semantic similarity.

LSA has been shown to reflect human knowledge in a variety of ways. For example, LSA measures correlate highly with humans’ scores on vocabulary tests, mimic human category judgments, predict how rapidly people are able to access words, and estimate passage coherence. In applied domains, LSA has been used to aid information retrieval, guide discussion forums, provide feedback to pilots on landing technique, diagnose mental disorders, select candidates for jobs, and allow automated tutors to understand the input they receive from students. The most surprising and controversial application of LSA has been its use in automated essay grading. Using the semantic vectors provided by LSA, it is possible to compare novel student essays to essays that have already been graded. If the new essay is most similar to the A essays it is awarded an A, and so forth. The accuracy of LSA at this task is remarkable. It has been consistently shown to correlate with human markers at rates equivalent to the agreement between humans.

...

  • Loading...
locked icon

Sign in to access this content

Get a 30 day FREE TRIAL

  • Watch videos from a variety of sources bringing classroom topics to life
  • Read modern, diverse business cases
  • Explore hundreds of books and reference titles

Read next

More like this

Sage Recommends

We found other relevant content for you on other Sage platforms.

close

Have you created a personal profile? Login or create a profile so that you can save clips, playlists and searches

close
close
close Signed In to profile You are signed in as:
close
Logged in Institution Institution