“Dongxu Zhang 14-11-03”版本间的差异

2014年11月2日 (日) 16:52的最后版本

Accomplished this week

• Create 100k,200k,150576 vocabulary. And use 150576 to build baiduhi, baiduzhidao language model(still running, preprocess).
• Use 166k vocabulary to train lm on baiduhi, baiduzhidao seperately,(still running ,pruning)
• Extract sentences which contains English and numbers from weibo corpus.
• Running BPTT using rwthlm. Still not normal. High ppl, low wer. But it seems that using rwthlm itself, lstm is indeed better than standard bptt.
• Found a tool called Shenlan which can parse Sogou cell vocabulary. Using its code with a crawler, we can update our vocabulary with new words.

Planned for next week

• Working on building lm and comparing vocabulary.
• Working on rwthlm.