Searching Semantic Memory as a Scale-Free Network: Evidence from Category Recall and a Wikipedia Model of Semantics

From Wikipedia Quality
Revision as of 15:43, 21 June 2020 by Liliana (talk | contribs) (infobox)
Jump to: navigation, search


Searching Semantic Memory as a Scale-Free Network: Evidence from Category Recall and a Wikipedia Model of Semantics
Authors
Graham William Thompson
Christopher T. Kello
Priscilla Montez
Publication date
2013
Links
Original

Searching Semantic Memory as a Scale-Free Network: Evidence from Category Recall and a Wikipedia Model of Semantics - scientific work related to Wikipedia quality published in 2013, written by Graham William Thompson, Christopher T. Kello and Priscilla Montez.

Overview

Searching Semantic Memory as a Scale-Free Network: Evidence from Category Recall and a Wikipedia Model of Semantics Graham William Thompson (gthompson2@ucmerced.edu) Christopher T. Kello (ckello@ucmerced.edu) Priscilla Montez (pmontez@ucmerced.edu) Cognitive and Information Sciences University of California, Merced 5200 North Lake Road, Merced, CA 95343 USA Abstract How is semantic memory structured and searched? Recalling items from semantic categories is a classic assay of semantic memory, and recall dynamics tend to exhibit semantic and temporal clustering, as if memory items are organized and retrieved in clusters. Recent analyses show this clustering to be approximately scale-free in terms of distributions of inter- retrieval intervals (IRIs). This finding is replicated and extended in the present study by asking participants to type as many animals as they can recall from semantic memory. To begin to explain these results, the organization of semantic memory is modeled as a network based on Wikipedia entries for nearly 6,000 animals. The Wikipedia animal network is found to be scale-free in terms of its degree distribution, and aspects of the network are found to correlate with aspects of recall. Semantic similarity based on Wikipedia entries is found to compare favorably with a measure based on latent semantic analysis. It is concluded that semantic memory processes can be usefully theorized as searches over scale- free networks. Keywords: semantic memory, scale-free networks, Levy foraging; category recall; latent semantic analysis; Wikipedia Introduction Category recall is a classic approach to investigating semantic memory. Participants produce as many items from a semantic category as possible in a specified period of time (Bousfield & Sedgewick, 1944). Items tend to be recalled in clusters. For the category of “animals”, for instance, part of a typical sequence might be “lion, tiger, cougar, leopard… kitten, cat, tabby”. This sequence contains two groups of semantically similar items, big wild cats followed by house cats. Such clusters can be of varying kinds and sizes, and they tend to correspond with short IRIs, relative to longer pauses when switching from one cluster to the next (Grunewald, Lockhead & Gregory, 1980). Clustering seems to be a general feature of semantic memory. Work in this area has a long history, with early experiments showing that, when participants memorize words presented in random order, they tend to recall those words in clusters based on semantic categories (Bousfield & Sedgewick, 1953). Therefore clustering must be related to memory encoding, retrieval, or both. In clinical work, semantic category recall is used as a diagnostic for mental disorders. Schizophrenic patients, people with semantic dementia and people with Alzheimer’s all show specific deficits in category generation tasks (see Murphy, Rich & Troyer, 2006). Previous work has sought to account for clustering in category recall with patch foraging models (see Hills, Jones & Todd, 2012). Patch foraging theorizes semantic memory as a set of patches of similar items. Memory search consists of series of quick retrievals of items from within a patch, interleaved by longer times for switching to the next patch when a sufficient number of items in the current patch have been found. Framed this way, optimal foraging can be expressed in terms of the time to leave a patch. It is optimal to switch when the instantaneous rate of recall per unit time drops below the long-term expected rate of recall (Charnov, 1976). Category switch times, and times in other human search tasks, have been found to be consistent with patch foraging (Cain, Vul & Mitroff, 2012). Patch foraging models lead one to expect short and long IRIs corresponding to successive recalls within patches versus between patches, respectively. However, recent work on category generation tasks has examined distributions of IRIs and found them to have no particular mean or means (Rhodes & Turvey, 2007). When the category recall task was conducted for sufficiently long periods of time (e.g. ten to twenty minutes for recalling animals), IRIs were found to be power law distributed. In particular, the frequency of IRIs were inversely related to their size, P(IRI) ~ 1/IRI α . Such power law distributions have no characteristic scale in theory, which means that their means and variances diverge as more samples are drawn. The implication is that the organization of semantic memory is scale-free rather than just patchy. Power law IRI distributions fall outside the purview of patch foraging models, but they have been studied extensively in animal foraging models (Viswanathan et al., 1996). Unlike patch models, animal foraging models explicitly consider the space in which items are to be found, such as trees and bushes in a meadow where birds are foraging for nuts and berries. Interestingly, the same power law distribution found in category recall is also found in IRIs during foraging for a wide range of species (Sims, Southall & Humphries, 2003). Theorists have related these findings to so-called Levy walks (Mandelbrot, 1982), which are random walks with path lengths drawn from a power law distribution. While it is unlikely that foraging paths are literally random Levy walks, they may capture an important property of foraging. The reason is that Levy walks sometimes may be efficient search strategies when their exponent α is near 2 (Viswanathan et al., 2000). Consistent