Estimating Semantic Networks of Groups and Individuals from Fluency Data

Abstract

One popular and classic theory of how the mind encodes knowledge is an associative semantic network, where concepts and associations between concepts correspond to nodes and edges, respectively. A major issue in semantic network research is that there is no consensus among researchers as to the best method for estimating the network of an individual or group. We propose a novel method (U-INVITE) for estimating semantic networks from semantic fluency data (listing items from a category) based on a censored random walk model of memory retrieval. We compare this method to several other methods in the literature for estimating networks from semantic fluency data. In simulations, we find that U-INVITE can recover semantic networks with low error rates given only a moderate amount of data. U-INVITE is the only known method derived from a psychologically plausible process model of memory retrieval and one of two known methods that we found to be consistent estimators of this process: if semantic memory retrieval is consistent with this process, the procedure will eventually estimate the true network (given enough data). We conduct the first exploration of different methods for estimating psychologically valid semantic networks by comparing people's similarity judgments of edges estimated by each network estimation method. To encourage best practices, we discuss the merits of each network estimation technique, provide a flow chart that assists with choosing an appropriate method, and supply code for others to employ these techniques on their own data.

Multiplex networks quantify robustness of the mental lexicon to catastrophic concept failures, aphasic degradation and ageing

“…Similarly, other approaches based on multiplex structure have been successful in predicting picture naming performance of people with aphasia 27 . Through a different network construction, structural network metrics were powerful in predicting also cognitive degratation due to Alzheimer's disease 14,33 . All these approaches indicate that investigating the network structure of the mental lexicon can be considered a first important step for understanding language processing, providing the foundations for analysing other important elements of the lexicon, such as search strategies, through additional research.…”

“…Such representation is used as a quantitative framework for exploring the large-scale robustness of the mental lexicon under progressive word failure through the formalism of network-based attacks 20,22,23,25 . Finding more or less efficient ways of disrupting connectivity within words in the mental lexicon can offer valuable insights about the cognitive organisation of words in the mind and, more importantly, also provide useful information aimed at understanding and then hampering or preventing cognitive decline in clinical pathologies where language is progressively disrupted, like aphasia 8,32 or Alzheimer's disease 14,33 .…”

SNAFU: The Semantic Network and Fluency Utility

“…Traditionally, the analysis of fluency data has been limited to behavioral measures such as the number of items listed or the number of cluster switches. More recently, fluency data has been used to estimate latent semantic networks of groups and individuals (Zemla & Austerweil, 2018). A semantic network is a representation consisting of a set of nodes (one for each word), and a set of edges that connect nodes that are semantically similar (e.g., horse and zebra may be connected by an edge).…”

“…As a result, a semantic network analysis of fluency data is rarely performed. Some network estimation methods are worse than others at capturing human behavior (Zemla & Austerweil, 2018), but choosing an estimation method is still ad hoc and often based on ease of implementation. Further, not having standards and best practices can lead to the temptation of selecting a network estimation method based on which one provides the desired results (as well as more innocuous forms of motivated data analysis).…”

“…SNAFU automates many common approaches to quantifying fluency data, including computing cluster sizes and switches, word frequencies, age-of-acquisition, intrusions, and perseverations. SNAFU also implements a number of methods for estimating networks from fluency data, and uses current best practices as defaults (Zemla & Austerweil, 2018). SNAFU comes in two flavors: (1) a Python library for programmatically analyzing fluency data, and (2) a graphical user interface (GUI) that provides an easy point-and-click interface for analyzing fluency data.…”