MIT LAB MTG MIT LAB MIT LAB MTG
MI&T LAB MTG推导举例
MI&T LAB 子图对
MI&T LAB 子图MTG系统框架 input source language sentence Bilingual parse tress Language model Training system Sub-graph pair table Decoding system output: bilingual “parse tree”
MI&T LAB 解码算法 1 Vector<rule> XRules; 2 for l=1 to Sent. Length 3 for i=0 to Sent. Length-l 4 j=l+i; 5 guessrule(i, j, XRules); 6 foreach rule in XRules 7 if rule exists in Graph table 8 add new rule to [X, i, j]; 9 Get k-best derivations of [X, i, j];
MI&T LAB 每个结点的k-best计算 Make a heap as a priority queue For every rule push its 1 st best translation into the heap While(candi. size()<k && heap. size() != 0) candi. push back(heap. top()) pop heap push neighbors of heap. top() into heap
MI&T LAB IWSLT 2004测试 System Pharaoh Sub-graph pairs BLEU-4 0. 3548± 0. 0082 0. 3714± 0. 0085
MI&T LAB 一些翻译结果 Pharaoh They must compensate our losses Subgraph They must compensate our losses Pharaoh They will be compensation for our losses Subgraph They will compensate our losses Pharaoh According to regulations, they will be our loss compensation Subgraph According to regulations, they will compensate our losses
MI&T LAB 谢 谢!
- Slides: 31