[ 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

Next revision
Previous revision
courses:rg:2011:deciphering_foreign_language [2011/12/06 09:52]
tran vytvořeno
courses:rg:2011:deciphering_foreign_language [2012/01/08 22:27] (current)
tran
Line 2: Line 2:
  
 Scriber: Ke. T Scriber: Ke. T
 +
 +The talk is about how to tackle MT without parallel training data.
 +
 +==== Section 1 ====
 +Given sentence pairs (e,f) where e is an English sentence and f is a foreign sentence, the translation model estimates parameter 
 +<latex>\theta</latex> such that
 +<latex>
 +\mathop {\arg \max }\limits_\theta  \prod\limits_\theta  {p_\theta  (f|e)} 
 +</latex>
 +
 +In case we do not have parallel data, we observe foreign text and try to maximize likelihood 
 +<latex>
 +\mathop {\arg \max }\limits_\theta  \prod\limits_f {p_\theta  (f)} 
 +</latex>
 +
 +Treating English translation as hidden alignment, our task is to find the parameter <latex>\theta</latex> that
 +<latex>
 +\mathop {\arg \max }\limits_\theta  \prod\limits_f {\sum\limits_e {P(e) \times \sum\limits_a {P_\theta  (f,a|e)} } } 
 +</latex>
 +
 +==== Section 2 ====
 +Section 2 deals with a simple version of translation, Word Substitution Decipherment, where there is only one-to-one mapping between source string and cipher string (the position of string does not change.)
 +
 +The solution for this problem is pretty simple: Given a sequence of English tokens <latex>e=e_1,e_2,...,e_n</latex>, and the corresponding sequence of cipher tokens <latex>c=c_1,c_2,...,c_n</latex>, we need to estimate parameter <latex>\theta</latex>:
 +<latex>
 +\mathop {\arg \max }\limits_\theta  \prod\limits_c {P_\theta  (c)}  = \mathop {\arg \max }\limits_\theta  \prod\limits_c {\sum\limits_e {P(e) \times \prod\limits_{i = 1}^n {P_\theta  (c_i |e_i )} } } 
 +</latex>
 +
 +The key idea of section 2 is the Iterative EM algorithm, which is used to estimate <latex>\theta</latex> more effectively in term of saving memory and running time (complexity).
 +
 +If we use traditional EM, every time we update <latex>\theta</latex>, we also need to update pseudo-counts (in this case, conditional probabilities <latex>{P_\theta  (c_i |e_i )</latex>.) It leads to O(|V|<sup>2</sup>) time. The heart of Iterative EM is that at every iteration, the algorithm run on a proportion of the most frequent words in vocabulary, and whenever the algorithm estimates <latex>P(c_i|e_i) > 0.5 </latex>, it fixes that probability equal to 1 in the following iteration, hence, the number of free parameters need to be estimated reduce after each iteration.
 +
 +__**Practical questions:**__ How to initiate EM? How to start the first iteration?
 +
 +**Some other notes related to this paper:**
 +  - Generative story: The generative process that generates data given some hidden variables.
 +  - [[http://en.wikipedia.org/wiki/Chinese_restaurant_process|Chinese restaurant process]]
 +  - Gibbs sampling:  an algorithm that generates a sequence of samples from the joint probability distribution of two or more random variables.
 +
 +Why did they experiment with Temporal expression corpus? This corpus has relatively small word types, it makes easier to compare Iterative EM with full EM.
 +
 +==== Section 3 ====
 +Not many details of this section was presented, however, there are few discussions around this.
 +
 +How to choose the best translation? After finishing parameter estimation, pick the final sample and extract the corresponding English translations for every for- eign sentence. This yields the final decipherment output.
 +
 +Given another text (which is not in training data), how to translate it? Use MLE to find the best translation from the model.
 +
 +==== Conclusion ====
 +This is an interesting paper, however, there is a lot of maths behind.  
 +
 +
  

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