In the CRF model, we model the sequence tagging problem with a global feature function and parameters w as follows: x = (x1, ..., xn) t = (t1, t2, ..., tn) p(t | x, w) = exp( w . Φ(x, t) ) / Σt' in Sn[ exp( w . Φ(x, t') ] 1. What are the two computational challenges that must be overcome to solve such a model? 2. What are the assumptions on the feature function Φ which are exploited to make CRF computationally tractable?
קורסי בחירה במדעי המחשב NLP2020
In the CRF model, we model the sequence tagging problem with a global feature function and parameters w as follows: x = (x1, ..., xn) t = (t1, t2, ..., tn) p(t | x, w) = exp( w . Φ(x, t) ) / Σt' in Sn[ exp( w . Φ(x, t') ] 1. What are the two computational challenges that must be overcome to solve such a model? 2. What are the assumptions on the feature function Φ which are exploited to make CRF computationally tractable?
1
| done | ||
מיין לפי
* השאלה נוספה בתאריך: 01-02-2020