|Busy with discrimination at a party. Image © | Dreamstime.com|
Imagine a cocktail party with so many people talking that you can barely recognise individual words any one of them is saying. Suddenly someone mentions your name, and that one piece immediately catches your attention. Why is that? Because your name means something to you? Of course it does, but so do many other words the people around you are using. (Otherwise you wouldn’t be at that party, right?)
Or imagine you just bought a new car. You immediately start noticing the same model on the street. Is it because everybody suddenly bought the same car to be like you? Not likely. A simpler explanation is that you started noticing something you hadn’t noticed before. You learned to discriminate a new subclass of objects in your environment, or distinguish this particular model from models similar to it.
This is called discrimination learning, a model of learning originating from classical conditioning (Pavlov) that has been successfully applied to language learning1, among other things. We distinguish between events if and only if that distinction somehow matters for us, if we have learned that some meaningful outcome depends on it.
In the cocktail party example above, the distinction between being talked about and not being talked about is way more important than distinguishing between any other topics those people at the next table might be discussing, which is why we have learned to pick out our name and other similarly relevant cues from the audio stream that our hearing system receives.
This works on other levels too. English native speakers distinguish between the vowels in “beer” and “bear” which many L2 English learners find hard, because they haven’t had enough exposure to learning material that would prove this distinction useful. Conversely, it is hard for English speakers to hear the difference between tones when learning a tonal language.
The relatedness graphs in qlaara are based on discriminative learning algorithms, mimicking the way humans learn to understand and produce speech. Search the qlaara dictionary for a word and see how the results of machine learning differ from your intuition.
1 For a short into, see e.g. Ramscar, M., & Baayen, H. (2013). Production, comprehension, and synthesis: a communicative perspective on language. Frontiers in Language Sciences, 233.↩