next up previous contents index
Next: Cache Up: Multilevel smoothing for trigram Previous: The full trigram model

Practical issues

Pooling of parameters

 

In smoothing trigram models by linear discounting  or interpolation , the smoothing parameters, both tex2html_wrap_inline45809 and tex2html_wrap_inline45811, have been assumed to depend on the word pair uv. The number of these word pairs itself is huge so that reducing the number of smoothing parameters is desirable. Often, the parameters are pooled or tied   across different histories by setting [Jelinek & Mercer (1980)]:
eqnarray9299
This means that the parameters are tied across histories h with the same count N(h). It is straightforward to repeat the derivations for this type of tying.  As result, we obtain:
eqnarray9304
Similarly, when assuming the parameters to be independent of the histories h, we obtain:
eqnarray9309
For absolute discounting , we obtain similar formulae in the case of tying.  In particular for absolute discounting , the experimental results show that there is no degradation in perplexity  when using history independent discounting parameters [Ney & Essen (1993)].  

Computational complexity

In a real implementation, no matter whether for off-line purposes or in an operational prototype  system, the computational complexity of the trigram or bigram  model has to be considered, namely the memory requirements and the access time.



next up previous contents index
Next: Cache Up: Multilevel smoothing for trigram Previous: The full trigram model

EAGLES SWLG SoftEdition, May 1997. Get the book...