Reginald Ferber
Projects & Areas of Interest

Print View ] [ DEUTSCH ]

 Projects and Areas of Interest
 +Information Retrieval
 +Digital Libraries
 +Associative Word Nets
 +Knowledge Discovery in Text Collections
 +eLearning Search and Broker System
 +Simulation of Word Recognition
 +Language and Gender
 +Cellular Automata
 Contact and Disclaimer

 +Cognitive Dynamics ...
 +Remark on Kawamoto's Approach
 +Korpusbasierte neuronale Netze
 +Simulation der Worterkennung

Simulation of Word Recognition with Associative Nets

The simulation of the selection of search terms with associative word nets was based on the assumption that associations between words or "smaller objects" play an important role in human knowledge representation and processing. I have further followed-up this line of arguments in a project funded by the DFG (German Research Foundation). Its goal was to simulate results of psychological experiments on the timing of word recognition and reading with an associative spreading activation net.

Experiments on Word Recognition and Reading

Recognition and reading of (single) words has been investigated in many psychological experiments. In the simplest case a word is presented on a screen to subjects and they have to read it aloud ("naming"). The time between the appearance of the word and the onset of reading is measured. Another task is the "lexical decision": a string of characters is presented to subjects on a screen and they have to press one button if the characters form a word and a different button of they don't. Again the time between the appearance of the letters on screen and the use of the button is measured. Time spans measured are in the range of up to one second and depend on the length and "familiarity" of the words or patterns presented.

Priming Experiments

In this more complex paradigm a different word or pattern of characters is displayed for a very short time before the actual target word is presented. This "prime" can be similar to the following target with regard to it's content or the pattern of characters. The point of interest is, how these similarities or dissimilarities influence the times for reading or lexical decision. Several studies have shown, that similar primes result in shorter reading or decision times.

The Simulation

To simulate the timing I used an "spreading-activation-net" whose nodes represented either words or (frequent) patterns of characters and had an "activation" value. The nodes where connected with weighted edges to spread the activation between nodes and compute new activation values for each node based on the activation of its neighbours. The weights of these connections where "trained" with a text corpus resulting in high weights for nodes that appeared frequently together in the text.

To simulate a word recognition experiment the node(s) of the target word have been activated and spreading cycles have been calculated until the configuration of values on the web did not change (or did change only slightly), i. e. until the web converged to a stable configuration (or until a maximum number of cycles was reached). The number of cycles was compared to the time measured in the experiments. To simulate priming experiments the nodes appearing in the prime were activated before the activation of the nodes of the target word. I have implemented the simulation program in C++. The basic ideas of the approach are presented in the conference paper Cognitive Dynamics - Dynamic Cognition? (Ferber 1996 [->]).


The simulations showed interesting results in the expected direction. There is - however - a lot room for further elaboration of the model, improvements of methods, and further experiments. An detailed report (in German) on the results is given in the technical report Simulation von Worterkennungsprozessen mit konnektionistischen Wortnetzen (Ferber & Wettler 1996 [->]). I was not able to follow up this exciting line of experiments, because German laws imposed a maximum employment time of five years in one university for scientist with non-tenure positions.

HTML file generated 25. 9. 2003 by R. Ferber