Arseny Tolmachev
Arseny Tolmachev
Actual ngram features depend on field entry pointers, so we can't change them. We can completely rebuild trie index and dictionary entry table. Logic should be: 1. Do the step...
Need to write a spec and mecab-unidic compatible output format.
And codegen is incorrect if there are several identical unigrams
Need to implement LSTM inference like a Mikolov RNN is implemented right now. We want to implement lightweight inference inside Juman++ directly because we want to do both batching and...
Most of the space is taken by embeddings and maxent model. We should at least quantize them. Probably, it's better to explore #53 before this.
GATE７二分おくれてる We should create 7 as a candidate.