P-Interest Meeting Today: Dozat
Join the P-Interest Group today as they hear from Timothy Dozat, who will be presenting on his computational phonology QP which models OT using neural networks. All are welcome!
Modeling OT constraints using Artificial Neural Networks
If one is to assume that OT is a plausible cognitive model of linguistic production and/or comprehension, then one must take a stance on whether constraint definitions are hardwired into humans’ brains from birth and must only be ranked, or inferred solely from the linguistic data learners are exposed to during acquisition, or some combination of the two. The strong position that all constraints are innate and the learner must only rank them is very difficult to support, suggesting that constraint definitions–as well as constraint rankings–must at least partially be learned. However, previous computational models attempting to show how constraint definitions can be learned from data have faced severe shortcomings, many stemming from the discrete nature of the the constraint definitions (e.g. assign a violation of weight w if features a and b are present in the input). I will show that allowing for continuous values in constraint definitions (e.g. assign p% of a violation of weight w if feature a is present in the input with weight v and feature b is present in the input with weight u) allows for constraints to be represented with artificial neural networks, which can make small changes to constraint definitions without radically changing their behavior or throwing them out entirely. This representation comes with all the perks of standard neural networks, to the effect that vowel harmony and constraint conjunction can be modeled with only small changes to the model.