improvement for accept generic arpa language-model data

[jg1uaa]

sortlm.py requires following rules:

• 2-gram use 1-gram word
• 3-gram use 2-gram word pair and 1-gram word

if this rule is not matched, sortlm.py stops with error.

data.arpa in libkkc-data is tuned to meet the conditons, but normally generated arpa files by LM utility (such as IRSTLM) are not.

so unknown word and word-pair should be discarded.

[jg1uaa]

I am trying to create data.arpa from Nihongo Web Corpus 2010 at https://github.com/jg1uaa/nwc2010-libkkc .