I'm interested in finding a solution for the following problem:
- problem space: any language that has more than 1 article
- let's take German language as example. So the articles in German are "der", "die", "das". - every word has an article as prefix and you must use the article. for example you can't say just "Auto"(car), you must say "das Auto".
- If you are not a native German speaker, you have bad luck, because you have to memorize the article too, when learning a new word, since there are only 3-4 rules to know which word takes which article. And those rules are perhaps about max. 10% percent of the vocabulary. (for example if a word ends with "-ung", it takes "die")
So here comes the funny part: as a non-native-German-speaker, i wanted to analyze the language from an IT-point of view and asked some friends of mine random words with the following properties:
- first I reduced the input set from "any German word" to "a German word", for which no known article rule exists".
- then I extended the input set with "made up words", which do not exist.
Every candidate had the same answer for German words, which should be not surprise. But when I asked them per word "why do you think/feel that the article of the word is "der/die/das"?" they couldnt give an answer. They just know it, without knowing why.
Here comes the real hammer: every candidate had the same answer for any made-up non-German word. I say anything, make any meaningful voice, and they all give the same answer (ie. article).
I'm pretty sure that I'm not the first human being in the world who made thoughts about this topic. And I'm interested in any scientific papers/research about that topic. By the way, what I definitely do NOT need is, any method (like neural networks) which outputs a pure mathematical function. I thought about following possibilities:
focusing on feelings of subjects psychologically (I think it can be managed with colors somehow etc.)
analyzing the words statistically according to their syllable structure
Where can I begin? What are your suggestions? Are there any Machine Learning research about that topic?