Scroll down for a description of the simulation and suggestions of things to try!
The applet requires Java 1.4.1 or higher. It will not run on Windows 95 or Mac OS 8 or 9. Mac users must have OS X 10.2.6 or higher and use a browser that supports Java 1.4. (Safari works, IE does not. Mac OS X comes with Safari. Open Safari and set it as your default web browser under Safari/Preferences/General.) On other operating systems, you may obtain the latest Java plugin from Sun's Java site.
created with NetLogo
view/download model file: LexicalCompetition.nlogo
This model specifically illustrates the effect of category competition on contrast within the context of (i) information gain through variation in use and (ii) information loss through memory decay. The broader purpose of the simulation is to illustrate how the lack of word-external disambiguation in context can promote retention of phonetic contrast in near-homophones.
As an example of word-external disambiguation, words like 'too' and 'two' in English are rarely confused by listeners because they are used in distinct sentential and semantic contexts. In contrast, words within morphological paradigms are often distinguished in context primarily by their phonetic differences. For example, utterances like 'Frank cooks chicken well' and 'Frank cooked chicken well' could be used in very similar contexts, in which case the present or past tense of the verb 'cook' is conveyed solely by the phonetic information present in the suffix.
The hypothesis explored in Blevins and Wedel (submitted) is that the greater 'functional load' of word-internal phonetic information within paradigms may account for anti-homophony effects in paradigms. 'Anti-homophony' refers to the failure of otherwise regular sound changes to occur in words when that change would render them homophonous with another word.
In this simulation, exemplars from two categories (red and blue) coexist in a 2-dimensional parameter space. In each round, the red category and blue category each produce an output. The output is an average of some number of randomly chosen exemplars from the category, plus some random noise. This noise can be biased slightly toward the center of the parameter space to model lenition. Depending on how you set up the variables in the simulation, the output may be stored as a new exemplar in the category that it came from, or it may be stored in the category it is closest to -- which turns out to make a big difference in how the simulation runs.
Storing the output in its originating category in this simulation models the case of word-external disambiguation in language, in which ambiguity in the form of a speaker-output is compensated by other information allowing the listener to assign it to the correct category despite homophony. In this case, lenition pushes the two categories to overlap.
Setting the simulation up so that category outputs are assigned to the closest category models the case in which the phonetic information in the speaker-output is the primary determining factor in category assignment by the listener. In this situation, ambiguous pronunciations are more likely to be assigned to the 'wrong' category than less ambiguous pronunciations. This variant trading along the boundary between the two categories as they evolve through new exemplar creation and old exemplar decay keeps the categories apart -- despite continuous lenition pushing them toward the center of the parameter space.
Start by clicking 'Setup'. The simulation starts out with 10 exemplars each of the red and blue categories, scattered at random around the parameter space.The red bullseye marks the category center for the red exemplars, and the blue bullseye marks the category center for the blue exemplars. As the category centers move, they leave a trace of the category movement behind them. Positions in the parameter space that are occupied by exemplars from both categorie are colored green. You can shift the setting of the 'show-exemplar' button at any time to make the exemplars invisible which better reveals the pathway of the two category centers over time.
The 'Go once' button allows you to step through the simulation one cycle at a time with each click, while the 'Go forever' button will, as you might guess, make the simulation go forever. To stop the simulation, simply click on 'Go forever' again.
The 'blending' slider allows you to control how many exemplars from a given category are averaged to produce an output from that category. The more exemplars are averaged, the closer the output will tend to be toward the category center.
The 'production-noise' slider adds some randomness to the position of the output. The more production noise, the farther the output will tend to be from the category center. As you can see, blending and production noise act antagonistically. The simulation works best when they are set similarly.
The 'memory decay' slider influences how fast exemplars are removed from the simulation. The lower the number, the slower the decay, the more exemplars there will be at any one time and the slower the categories will evolve. Since exemplars can be on top of one another, you may not be able to immediately notice that more exemplars exist when you reduce memory decay.
The 'lenition' slider lets you set the degree of bias in production toward the center of the field. A setting of '1' introduces a 1% bias in random noise toward the center of the field.
Finally, the 'category competition' slider lets you set the percentage of the time that a new output's category assignment is dependent on its relative closeness to the two categories. When the slider is set at 0, new outputs are assigned to the category they came from. When the slider is set at 50, for example, half of the time they are assigned to their originating category without reference to their value, and half of the time they are assigned to the closest category, whatever that is.
Note that the simulation runs faster when you have 'category competition' set at a lower number. For purposes of comparison across different runs, the number of cycles ('ticks') is recorded at the top of the field.
Start out with the sliders in their 'setup' positions, that is:
blending = 5
production noise = 5
memory decay = 30
lenition =
1.5
category competition = 0
Now click on 'Go forever'. Adjust the speed slider to get the speed you like.
You should see the initially random spread of exemplars rapidly coalesce into tight distributions, which then randomly drift around the field. You'll soon notice that they overlap and cross one another frequently.
Stop the simulation by clicking on 'Go forever' again, and then click 'set up' again to reset the simulation. Now set 'Category competition' at or near 100, and start the simulation again. This time, you'll notice that while the two categories sometimes approach one another, they do remain separate in space. You will be able to see the pathway of each category center over time better with the 'show-exemplar' button set to 'off'. A good comparison can be made by setting the 'show-exemplar' button to 'off', and running the simulation two times out to 50-100 ticks once with 'category-competition' set to 0, and once with it set to >50.
How little category competition can you get away with to keep categories apart? In this implementation, at least, 50% category competition seems to be plenty. Under this model then, two near-homophones do not need to be always solely distinguished in usage by their phonetic contents for homophony-creating sound change to be blocked -- just some of the time.