[ 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

Both sides previous revision Previous revision
courses:rg:2012:jodaiberreport [2012/03/26 18:45]
joda
courses:rg:2012:jodaiberreport [2012/03/26 18:51] (current)
joda
Line 43: Line 43:
 </ul> </ul>
 <h2 id="predicitive-exchange-clustering">Predicitive Exchange Clustering</h2> <h2 id="predicitive-exchange-clustering">Predicitive Exchange Clustering</h2>
-<p><br /><span class="math">+<p>(for formula see PDF version or paper)</p>
-    P({w_i}|w_1^{i - 1}\approx {p_0}({w_i}|c({w_i})) \cdot {p_1}(c({w_i}|{w_{i - 1}})) = \frac{{N({w_i})}}{{N(c({w_i}))}} \cdot \frac{{N({w_{i - 1}},c({w_i})}}{{N({w_{i - 1}})}} +
-    $</span><br /></p>+
 <ul style="color: black;"> <ul style="color: black;">
 <li>equations (6)-(10) in the paper demonstrate the new way to calculate the perplexity: when moving a word from c to c’, last part of (10) ( <span class="math"> - ∑ <sub><em>c</em> ∈ <em>C</em></sub><em>N</em>(<em>c</em>) ⋅ log<em>N</em>(<em>c</em>)</span> ) must be recalculated, the first part ( <span class="math">∑ <sub><em>v</em> ∈ <em>V</em>, <em>c</em> ∈ <em>s</em><em>u</em><em>c</em>(<em>v</em>)</sub><em>N</em>(<em>v</em>, <em>c</em>) ⋅ log<em>N</em>(<em>v</em>, <em>c</em>)</span> ) can be quickly calculated with an additional array</li> <li>equations (6)-(10) in the paper demonstrate the new way to calculate the perplexity: when moving a word from c to c’, last part of (10) ( <span class="math"> - ∑ <sub><em>c</em> ∈ <em>C</em></sub><em>N</em>(<em>c</em>) ⋅ log<em>N</em>(<em>c</em>)</span> ) must be recalculated, the first part ( <span class="math">∑ <sub><em>v</em> ∈ <em>V</em>, <em>c</em> ∈ <em>s</em><em>u</em><em>c</em>(<em>v</em>)</sub><em>N</em>(<em>v</em>, <em>c</em>) ⋅ log<em>N</em>(<em>v</em>, <em>c</em>)</span> ) can be quickly calculated with an additional array</li>

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