<div dir="ltr"><br>Statistical Machine Translation relies on parallel corpora (eg.. europarl) for training translation models. However these corpora are limited and take time to create. Yalign is designed to automate this process by finding sentences that are close translation matches from comparable corpora. This opens up avenues for harvesting parallel corpora from sources like translated documents and the web.<br>
<br><br>Project home: <a href="http://yalign.machinalis.com">http://yalign.machinalis.com</a><br>Code: <a href="http://www.github.com/machinalis/yalign">http://www.github.com/machinalis/yalign</a><br>Give Yalign a try and align some text extracted from Freebase: <a href="http://yalign.machinalis.com/demo">http://yalign.machinalis.com/demo</a><br>
<br>Yalign is implemented using:<br><br>+ A sentence similarity metric. Given two sentences it produces a rough estimate (a number between 0 and 1) of how likely are those two sentences to be a translation of each other.<br>
+ A sequece aligner, such that given two documents (a list of sentences) it produces an alignment which maximizes the sum of the individual (per sentence pair) similarities.<br>So Yalign’s main algorithm is actually a pretty wrapper to a standard sequence alignment algorithm.<br>
<br>For the sequence alignment Yalign uses a variation of the Needleman-Wunch algorithm to find an optimal alignment between the sentences in two given documents. On the good side, the algorithm has polynomial time worst case complexity and it produces an optimal alignment. On the bad side it can’t handle alignments that cross each other or alignments from two sentences into a single one (even tough is possible to modify the current implementation to handle those cases).<br>
<br>Since the sentence similarity is a computationally expensive operation, the mentioned “variation” on the Needleman-Wunch algorithm consists in using the A* to explore the search space instead of using the classical dynamic programming aproach (which would always requiere N * M calls to the sentence similarity metric).<br>
<br>After the alignment, only sentences that have a high probability of being translations are included in the final alignment. Ie, the result is filtered in order to deliver high quality alignments. To do this, a threshold value is used such that if the sentence similarity metric is bad enough that pair is excluded.<br>
<br>For the sentence similarity metric the algorithm uses a statistical classifier’s likelihood output and adapts it into the 0-1 range.<br><br>The classifier is trained to determine if a pair of sentences are translations of each other or not (a binary value). The particular classifier used for this project is a Support Vector Machine. Besides being excelent classifiers, SVMs can provide a distance to the separation hyperplane during classification, and this distance can be easily modified using a Sigmoid Function to return a likelihood between 0 and 1.<br>
<br>The use of a classifier means that the quality of the alignment is dependent not only on the input but also on the quality of the trained classifier.<br><br><br>Join our mailing list: <a href="http://groups.google.com/group/yalign">http://groups.google.com/group/yalign</a><br>
<br>See you soon!<br><br>floss [at] machinalis [dot] com<br><div><br></div>-- <br><span><font color="#888888">Elías Andrawos<br>Machinalis - <a href="http://machinalis.com/" target="_blank">http://machinalis.com/</a> <br>
+54 (0351) 152-750975</font></span><div><span><font color="#888888">+54 (0351) 4315739</font></span><span><font color="#888888"><br>Skype: eandrawos<br>Gtalk: <a href="mailto:eandrawos@machinalis.com" target="_blank">eandrawos@machinalis.com</a></font></span></div>
</div>