Maschineller Spracherwerb. Verstärkungslernen minimalistischer Grammatiken

We are developing a reinforcement learning algorithm for numeral words of natural languages. Numerals have an obvious connection between syntax and semantics, e.g. "two hundred and fourty-five" means 2 * 100 + 40 + 5. Our algorithm is intended to not only learn words but also to predict their integer meaning.

State-of-the-art algorithms try this by slot-filling of digit, i.e. they learn how to decide that 'two' is the hundred, 'four' is the ten… This requires a high degree of individual supervision for single languages – especially if their number systems are not even decimal. Instead, we use more general constraints of numerals word structure that are derived from Hurford's theory of the Packing Strategy [Hur06].
Our algorithm forms new words via exchange of subnumerals and their meaning is predicted by linear regression, see figure.

The exchange of subnumerals works by generalizing expressions like "two hundred and fourty-five" and converting them to an output set of words (X,Y) |--> "X hundred and Y".
In order to decide which subnumerals become generalized and which do not, we use general size and divisibility arguments. We represent the numeral systems by a minimalist grammar (MG).

Moreover, we want to investigate the possibility to utilize search engines as supervisors: Running the wrong word “fiveteen” in a search engine will display significantly less search results opposed to a search of the correct word “sixteen”. Based on that, we attempt to find a model to decide whether or not a word is correct.

• Natural numeral grammars can be efficiently represented as Minimalist Grammars
• Those Minimalist Grammars can be efficiently generated based on a learning algorithm that
  • reinforces itself by linear regression and
  • is supervised by big search engines