A short review of Multilink’s components
This section provides more detailed theoretical information on aspects of Multilink.
The graphic of the lexical network below illustrates the different representations in the network and their connections. Orthographic, phonological, semantic, and language representations exist for words in both Dutch and English. An input word activates orthographic representations depending on their similarity (Levenshtein distance). Next, information flows through the network, activating representations that are linked to orthography. Depending on the task at hand, an output response is given when certain criteria are fulfilled.
Multilink’s lexicon currently contains 1466 pairs of 3-8 letter Dutch words and their most prominent 3-8 letter English translation equivalents. For both Dutch and English words, frequencies of usage (in occurrences per million) and phonological specifications (SAMPA) are included.
The frequencies in Occurrences Per Million (OPM) are largely derived from SUBTLEX-UK and SUBTLEX-NL; in a few cases a CELEX frequency was chosen when other information was missing. Note that all English frequencies have been divided by 4, in order to simulate the performance of unbalanced Dutch (L1) - English (L2) bilinguals. We are currently evaluating this admittedly arbitrary setting for subjective frequency.
For demonstration purposes, the lexicon contains some interlingual homographs (defined here as two word forms from different languages with the same orthography but a different meaning) and cognates (translation equivalents with form overlap). However, the set of currently included homographs is not exhaustive or representative for the languages as a whole.
Here is an example of several lines from the lexicon:
In Multilink, a semantic network can be activated to simulate spreading activation in semantic priming tasks. The connection strengths between concepts can be derived from association databases like the English Miami corpus by Nelson & McEvoy at and the Dutch Word Association Database from DeDeyne & Storms. At present, Multilink’s parameter settings are not fit for simulating semantic priming with any degree of precision.
Multilink assumes that the activation spreading through the network can be used differently depending on task demands and stimulus list composition. To simulate different read-outs and task schemas, task-dependent decision criteria are implemented. At present, the task implementation is very simple: Recognition and lexical decision tasks read out orthographic activation, word naming and word translation read out phonology, semantic priming reads out semantics, and language decision reads out language membership information.
Several aspects of Multilink can be seen as fixated components of the architecture. For instance, there are minimum and maximum values of the activation, there is a fixed relationship between word frequency and resting level activation (RLA), and connection strengths between different codes (orthography, phonology, semantics, and language nodes) are fixated. The simulations of all studies discussed in papers have been conducted with only one setting of the parameters!
Log word frequency is used to set the resting level activation of words to a value between 0 and -0.2. It can be shown mathematically that this leads to a good fit with lexical decision data (Thul, Conklin, & Dijkstra, in preparation).
Orthographic and phonological representations of words compete during processing. This competition is implemented in terms of lateral inhibition (Van Geffen, Hieselaar, & Dijkstra, in preparation). This is different from response competition, which can be seen as taking place at the decision level. Current simulations suggest that in order to account for the performance of interlingual homographs in some task contexts (e.g. Vanlangendonck, Peeters, Rueschemeyer, & Dijkstra, 2019), both lateral inhibition and response competition are required.
Semantics and semantic priming
At present, Multilink implements a very simple semantics: Every word has its own concept. Cognates and translation equivalents are assumed to be mapped onto the same concept, which works fine. To simulate semantic priming effects (e.g.,
0:LAW,5:ORDER), an association database is currently used. It turns out this leads to several problems. To clarify what is going on and improve the simulations, we are currently investigating homonymy and polysemy from different perspectives: considering concepts as symbolic representations, distributed feature sets, or vectors.