“NLP Status Report 2017-7-31”版本间的差异

• made the poster for ACL
• attempted to fix repeated word, but failed
• done some work of n-gram model of the couplet

• generate streame according to a couplet
• complete the task of filling in the blanks of a couplet

• Got 55,000+ Englsih poems and 260,000+ lines after preprocessing
• Added phase separators as the style indicator, and every line has at least one separator
• Training loss didn't decrease very much, only from 440 to 50
• The translation quality deteriorated when added language model

• Try to use a larger language model to decrease the training loss
• Try to use character-based MT in English-Chinese translation

• looked for the performance(the bleu value) of other models

 on the WMT2014 dataset from the published papers,but not found.

• installed and built Moses on the server

• train statistical machine translation model and test it

 toolkit: Moses
 data sets:WMT2014 en-de、en-fr data sets

• collate experimental results.compare our baseline model with Moses

• process document.
• Shiji has been split up to 2,5000 pairs of sentence.
• Zizhitongjian has been split up to 2,0000 pairs.

• adjust jieba source code, in order to make jieba more accurate for ancient language wordpiece
• read model source code