Before this commit, the Mecab tokenizer had to be called twice when
creating a Doc- once during tokenization and once during tagging. This
creates a JapaneseDoc wrapper class for Doc that stashes the parsed
tokenizer output to remove redundant processing. -POLM
Added functionality for spans and docs to get lowest common ancestor
matrix by simply calling: doc.get_lca_matrix() or
doc[:3].get_lca_matrix().
Corresponding unit tests were also added under spacy/tests/doc and
spacy/tests/spans.
Designed to address: https://github.com/explosion/spaCy/issues/969.
Tagger is now parametrized and has two sentences with more tag coverage.
The tokenizer tests are updated to reflect differences in tokenization
between IPAdic and Unidic. -POLM
Example of this is いる in these sentences:
彼はそこにいる。# should be VERB
彼は底に立っている。# should be AUX
Unclear which case is more numerous - need to check a large corpus - but
in keeping with the other ambiguous tags, this is mapped to the
"dominant" or first part of the tag. -POLM
This is far from complete but it should be enough to check some things.
1. Mecab transition. Janome doesn't support Unidic, only IPAdic, but UD
tag mappings are based on Unidic. This switches out Mecab for Janome to
get around that.
2. Raw tag extension. A simple tag map can't meet the specifications for
UD tag mappings, so this adds an extra field to ambiguous cases. For
this demo it just deals with the simplest case, which only needs to look
at the literal token. (In reality it may be necessary to look at the
whole sentence, but that's another issue.)
3. General code structure. Seems nobody else has implemented a custom
Tagger yet, so still not sure this is the correct way to pass the
vocabulary around, for example.
Any feedback would be greatly appreciated. -POLM
The tokenizer caches output for common chunks, for efficiency. This
cache is be invalidated when the tokenizer rules change, e.g. when a new
special-case rule is introduced. That's what was causing #1061.
When the cache is flushed, we free the intermediate token chunks.
I *think* this is safe --- but if we start getting segfaults, this patch
is to blame. The resolution would be to simply not free those bits of
memory. They'll be freed when the tokenizer exits anyway.
