[ Skip to the content ]

Institute of Formal and Applied Linguistics Wiki


[ Back to the navigation ]

Differences

This shows you the differences between two versions of the page.

Link to this comparison view

courses:rg:2012:searn-in-practice [2012/09/25 02:37]
galuscakova vytvořeno
courses:rg:2012:searn-in-practice [2012/09/25 14:48] (current)
popel some answers to the questions
Line 19: Line 19:
 ===== Questions ===== ===== Questions =====
   *  It seems to me that we need ground true outputs for the testing data to run the algorithm (otherwise we cannot get the optimal policy), what makes no sense to me.   *  It seems to me that we need ground true outputs for the testing data to run the algorithm (otherwise we cannot get the optimal policy), what makes no sense to me.
 +     * No, we need "ground true" labels during training for the initial (optimal) policy. Each iteration interpolates two policies. However, this does not mean interpolating weight vectors (or the trained models in general). It means we have stored both of the policies and during decoding (testing) we make a coin flip and with probability beta we choose one of the policies. In the Nth iteration we are effectively interpolating N policies (the first is the optimal one). During testing we use just the last policy (so no need for "ground true" during testing).
   * Similarly and probably based on something that I missed: why do we want to move away from optimal policy completely. Maybe it is because at the end of the algorithm we return the current policy without the optimal policy. But what does "without" mean in this case?   * Similarly and probably based on something that I missed: why do we want to move away from optimal policy completely. Maybe it is because at the end of the algorithm we return the current policy without the optimal policy. But what does "without" mean in this case?
   * I do not see why the first condition in formula (4) is y<sub>t</sub> ∈ {begin X, out}. Shouldn't it be y<sub>t</sub> ∈ {out}?   * I do not see why the first condition in formula (4) is y<sub>t</sub> ∈ {begin X, out}. Shouldn't it be y<sub>t</sub> ∈ {out}?
 +    * No. We count the percentage of words which are correctly tagged as named entities, not the percentage of named entities.
   * Are there any real world applications using Searn?   * Are there any real world applications using Searn?
   * Are there any kind of problems, in which could Searn especially help?   * Are there any kind of problems, in which could Searn especially help?

[ Back to the navigation ] [ Back to the content ]