Bloom-filter backed Lookup Tables (#4268)

* Improve load_language_data helper

* WIP: Add Lookups implementation

* Start moving lemma data over to JSON

* WIP: move data over for more languages

* Convert more languages

* Fix lemmatizer fixtures in tests

* Finish conversion

* Auto-format JSON files

* Fix test for now

* Make sure tables are stored on instance

* Update docstrings

* Update docstrings and errors

* Update test

* Add Lookups.__len__

* Add serialization methods

* Add Lookups.remove_table

* Use msgpack for serialization to disk

* Fix file exists check

* Try using OrderedDict for everything

* Update .flake8 [ci skip]

* Try fixing serialization

* Update test_lookups.py

* Update test_serialize_vocab_strings.py

* Lookups / Tables now work

This implements the stubs in the Lookups/Table classes. Currently this
is in Cython but with no type declarations, so that could be improved.

* Add lookups to setup.py

* Actually add lookups pyx

The previous commit added the old py file...

* Lookups work-in-progress

* Move from pyx back to py

* Add string based lookups, fix serialization

* Update tests, language/lemmatizer to work with string lookups

There are some outstanding issues here:

- a pickling-related test fails due to the bloom filter
- some custom lemmatizers (fr/nl at least) have issues

More generally, there's a question of how to deal with the case where
you have a string but want to use the lookup table. Currently the table
allows access by string or id, but that's getting pretty awkward.

* Change lemmatizer lookup method to pass (orth, string)

* Fix token lookup

* Fix French lookup

* Fix lt lemmatizer test

* Fix Dutch lemmatizer

* Fix lemmatizer lookup test

This was using a normal dict instead of a Table, so checks for the
string instead of an integer key failed.

* Make uk/nl/ru lemmatizer lookup methods consistent

The mentioned tokenizers all have their own implementation of the
`lookup` method, which accesses a `Lookups` table. The way that was
called in `token.pyx` was changed so this should be updated to have the
same arguments as `lookup` in `lemmatizer.py` (specificially (orth/id,
string)).

Prior to this change tests weren't failing, but there would probably be
issues with normal use of a model. More tests should proably be added.

Additionally, the language-specific `lookup` implementations seem like
they might not be needed, since they handle things like lower-casing
that aren't actually language specific.

* Make recently added Greek method compatible

* Remove redundant class/method

Leftovers from a merge not cleaned up adequately.

spacy/lang/el/lemmatizer/__init__.py

@@ -46,9 +46,9 @@ class GreekLemmatizer(object):
         )
         return lemmas
 
-    def lookup(self, string):
-        if string in self.lookup_table:
-            return self.lookup_table[string]
+    def lookup(self, orth, string):
+        if orth in self.lookup_table:
+            return self.lookup_table[orth]
         return string
 
 

spacy/lang/fr/lemmatizer/__init__.py

@@ -52,7 +52,7 @@ class FrenchLemmatizer(object):
         elif univ_pos in (SCONJ, "SCONJ", "sconj"):
             univ_pos = "sconj"
         else:
-            return [self.lookup(string)]
+            return [self.lookup(None, string)]
         # See Issue #435 for example of where this logic is requied.
         if self.is_base_form(univ_pos, morphology):
             return list(set([string.lower()]))
@@ -114,9 +114,9 @@ class FrenchLemmatizer(object):
     def punct(self, string, morphology=None):
         return self(string, "punct", morphology)
 
-    def lookup(self, string):
-        if string in self.lookup_table:
-            return self.lookup_table[string][0]
+    def lookup(self, orth, string):
+        if orth is not None and orth in self.lookup_table:
+            return self.lookup_table[orth][0]
         return string
 
 

spacy/lang/nl/lemmatizer/__init__.py

@@ -62,11 +62,11 @@ class DutchLemmatizer(object):
             # are not lemmatized. They are lowercased, however.
             return [string]
             # if string in self.lemma_index.get(univ_pos)
-        lemma_index = self.index.get(univ_pos, {})
+        lemma_index = self.index.get_string(univ_pos, {})
         # string is already lemma
         if string in lemma_index:
             return [string]
-        exceptions = self.exc.get(univ_pos, {})
+        exceptions = self.exc.get_string(univ_pos, {})
         # string is irregular token contained in exceptions index.
         try:
             lemma = exceptions[string]
@@ -75,12 +75,12 @@ class DutchLemmatizer(object):
             pass
         # string corresponds to key in lookup table
         lookup_table = self.lookup_table
-        looked_up_lemma = lookup_table.get(string)
+        looked_up_lemma = lookup_table.get_string(string)
         if looked_up_lemma and looked_up_lemma in lemma_index:
             return [looked_up_lemma]
 
         forms, is_known = lemmatize(
-            string, lemma_index, exceptions, self.rules.get(univ_pos, [])
+            string, lemma_index, exceptions, self.rules.get_string(univ_pos, [])
         )
 
         # Back-off through remaining return value candidates.
@@ -103,9 +103,12 @@ class DutchLemmatizer(object):
     # Overrides parent method so that a lowercased version of the string is
     # used to search the lookup table. This is necessary because our lookup
     # table consists entirely of lowercase keys.
-    def lookup(self, string):
+    def lookup(self, orth, string):
         string = string.lower()
-        return self.lookup_table.get(string, string)
+        if orth is not None:
+            return self.lookup_table.get(orth, string)
+        else:
+            return self.lookup_table.get_string(string, string)
 
     def noun(self, string, morphology=None):
         return self(string, "noun", morphology)

spacy/lang/ru/lemmatizer.py

@@ -115,7 +115,7 @@ class RussianLemmatizer(Lemmatizer):
     def pron(self, string, morphology=None):
         return self(string, "pron", morphology)
 
-    def lookup(self, string):
+    def lookup(self, orth, string):
         analyses = self._morph.parse(string)
         if len(analyses) == 1:
             return analyses[0].normal_form

spacy/lang/uk/lemmatizer.py

@@ -112,7 +112,7 @@ class UkrainianLemmatizer(Lemmatizer):
     def pron(self, string, morphology=None):
         return self(string, "pron", morphology)
 
-    def lookup(self, string):
+    def lookup(self, orth, string):
         analyses = self._morph.parse(string)
         if len(analyses) == 1:
             return analyses[0].normal_form

spacy/language.py

@@ -32,6 +32,7 @@ from .lang.tokenizer_exceptions import TOKEN_MATCH
 from .lang.tag_map import TAG_MAP
 from .lang.lex_attrs import LEX_ATTRS, is_stop
 from .errors import Errors, Warnings, deprecation_warning
+from .strings import hash_string
 from . import util
 from . import about
 

spacy/lemmatizer.py

@@ -93,9 +93,9 @@ class Lemmatizer(object):
     def punct(self, string, morphology=None):
         return self(string, "punct", morphology)
 
-    def lookup(self, string):
-        if string in self.lookup_table:
-            return self.lookup_table[string]
+    def lookup(self, orth, string):
+        if orth in self.lookup_table:
+            return self.lookup_table[orth]
         return string
 
 

spacy/lookups.py

@@ -1,4 +1,4 @@
-# coding: utf8
+# coding: utf-8
 from __future__ import unicode_literals
 
 import srsly
@@ -6,7 +6,12 @@ from collections import OrderedDict
 
 from .errors import Errors
 from .util import SimpleFrozenDict, ensure_path
+from .strings import hash_string
 
+from . import util
+
+import srsly
+from preshed.bloom import BloomFilter
 
 class Lookups(object):
     """Container for large lookup tables and dictionaries, e.g. lemmatization
@@ -14,10 +19,6 @@ class Lookups(object):
     so they can be accessed before the pipeline components are applied (e.g.
     in the tokenizer and lemmatizer), as well as within the pipeline components
     via doc.vocab.lookups.
-
-    Important note: At the moment, this class only performs a very basic
-    dictionary lookup. We're planning to replace this with a more efficient
-    implementation. See #3971 for details.
     """
 
     def __init__(self):
@@ -54,8 +55,7 @@ class Lookups(object):
         """
         if name in self.tables:
             raise ValueError(Errors.E158.format(name=name))
-        table = Table(name=name)
-        table.update(data)
+        table = Table(name=name, data=data)
         self._tables[name] = table
         return table
 
@@ -100,10 +100,9 @@ class Lookups(object):
         bytes_data (bytes): The data to load.
         RETURNS (Lookups): The loaded Lookups.
         """
-        self._tables = OrderedDict()
-        msg = srsly.msgpack_loads(bytes_data)
-        for key, value in msg.items():
-            self._tables[key] = Table.from_dict(value)
+        for key, value in srsly.msgpack_loads(bytes_data).items():
+            self._tables[key] = Table(key)
+            self._tables[key].update_raw(value)
         return self
 
     def to_disk(self, path, **kwargs):
@@ -137,8 +136,10 @@ class Lookups(object):
 
 
 class Table(OrderedDict):
-    """A table in the lookups. Subclass of OrderedDict that implements a
+    """A table in the lookups. Subclass of builtin dict that implements a
     slightly more consistent and unified API. 
+
+    Includes a Bloom filter to speed up missed lookups.
     """
 
     @classmethod
@@ -153,15 +154,81 @@ class Table(OrderedDict):
         self.update(data)
         return self
 
-    def __init__(self, name=None):
+    def __init__(self, name=None, data=None):
         """Initialize a new table.
 
         name (unicode): Optional table name for reference.
+        data (dict): Initial data, used to hint Bloom Filter.
         RETURNS (Table): The newly created object.
         """
         OrderedDict.__init__(self)
         self.name = name
+        # assume a default size of 1M items
+        size = 1E6
+        if data and len(data) > 0:
+            size = len(data)
+
+        self.bloom = BloomFilter.from_error_rate(size)
+
+        if data:
+            self.update(data)
 
     def set(self, key, value): 
-        """Set new key/value pair. Same as table[key] = value."""
+        """Set new key/value pair, where key is an integer. Same as
+        table[key] = value.
+        """
         self[key] = value
+
+    def __setitem__(self, key, value):
+        OrderedDict.__setitem__(self, key, value)
+        self.bloom.add(key)
+
+    def set_string(self, key, value):
+        """Set new key/value pair, where key is a string to be hashed.
+        """
+        hkey = hash_string(key)
+        self.set(hkey, value)
+
+    def update(self, data):
+        """Add entries in a dict-like to the table, where keys are strings to
+        be hashed.
+        """
+        for key, val in data.items():
+            self.set_string(key, val)
+
+    def update_raw(self, data):
+        """Add entries in a dict-like to the table, where keys are ints.
+        """
+        for key, val in data.items():
+            self.set(key, val)
+
+    def get(self, key, default=None):
+        return OrderedDict.get(self, key, default)
+
+    def get_string(self, key, default=None):
+        hkey = hash_string(key)
+        return OrderedDict.get(self, hkey, default)
+
+    def __contains__(self, key):
+        # This can give a false positive, so we need to check it after
+        if key not in self.bloom: 
+            return False
+        return OrderedDict.__contains__(self, key)
+
+    def contains_string(self, key):
+        hkey = hash_string(key)
+        return self.__contains__(hkey)
+
+    def to_bytes(self):
+        # TODO: serialize bloom too. For now just reconstruct it.
+        return srsly.msgpack_dumps({'name': self.name, 'dict': dict(self.items())})
+
+    def from_bytes(self, data):
+        loaded = srsly.msgpack_loads(data)
+        self.name = loaded['name']
+        for key, val in loaded['dict'].items():
+            self[key] = val
+            self.bloom.add(key)
+
+        return self
+

spacy/morphology.pyx

@@ -273,7 +273,7 @@ cdef class Morphology:
         """
         if token.lemma == 0:
             orth_str = self.strings[token.lex.orth]
-            lemma = self.lemmatizer.lookup(orth_str)
+            lemma = self.lemmatizer.lookup(token.lex.orth, orth_str)
             token.lemma = self.strings.add(lemma)
 
     cdef int assign_tag(self, TokenC* token, tag_str) except -1:

spacy/tests/doc/test_creation.py

@@ -5,11 +5,13 @@ import pytest
 from spacy.vocab import Vocab
 from spacy.tokens import Doc
 from spacy.lemmatizer import Lemmatizer
+from spacy.lookups import Table
 
 
 @pytest.fixture
 def lemmatizer():
-    return Lemmatizer(lookup={"dogs": "dog", "boxen": "box", "mice": "mouse"})
+    lookup = Table(data={"dogs": "dog", "boxen": "box", "mice": "mouse"})
+    return Lemmatizer(lookup=lookup)
 
 
 @pytest.fixture

spacy/tests/lang/lt/test_lemmatizer.py

@@ -17,4 +17,4 @@ TEST_CASES = [
 
 @pytest.mark.parametrize("tokens,lemmas", TEST_CASES)
 def test_lt_lemmatizer(lt_lemmatizer, tokens, lemmas):
-    assert lemmas == [lt_lemmatizer.lookup(token) for token in tokens]
+    assert lemmas == [lt_lemmatizer.lookup_table.get_string(token, token) for token in tokens]

spacy/tests/lang/nl/test_lemmatizer.py

@@ -133,11 +133,11 @@ def test_nl_lemmatizer_pronoun_lemmas(nl_lemmatizer, text, lemma):
 # Using the lemma lookup table only
 @pytest.mark.parametrize("text,lemma", noun_irreg_lemmatization_cases)
 def test_nl_lemmatizer_lookup_noun(nl_lemmatizer, text, lemma):
-    lemma_pred = nl_lemmatizer.lookup(text)
+    lemma_pred = nl_lemmatizer.lookup(None, text)
     assert lemma_pred in (lemma, text)
 
 
 @pytest.mark.parametrize("text,lemma", verb_irreg_lemmatization_cases)
 def test_nl_lemmatizer_lookup_verb(nl_lemmatizer, text, lemma):
-    lemma_pred = nl_lemmatizer.lookup(text)
+    lemma_pred = nl_lemmatizer.lookup(None, text)
     assert lemma_pred in (lemma, text)

spacy/tests/vocab_vectors/test_lookups.py

@@ -19,9 +19,9 @@ def test_lookups_api():
     table = lookups.get_table(table_name)
     assert table.name == table_name
     assert len(table) == 2
-    assert table.get("hello") == "world"
-    table.set("a", "b")
-    assert table.get("a") == "b"
+    assert table.get_string("hello") == "world"
+    table.set_string("a", "b")
+    assert table.get_string("a") == "b"
     table = lookups.get_table(table_name)
     assert len(table) == 3
     with pytest.raises(KeyError):
@@ -50,10 +50,10 @@ def test_lookups_to_from_bytes():
     assert "table2" in new_lookups
     table1 = new_lookups.get_table("table1")
     assert len(table1) == 2
-    assert table1.get("foo") == "bar"
+    assert table1.get_string("foo") == "bar"
     table2 = new_lookups.get_table("table2")
     assert len(table2) == 3
-    assert table2.get("b") == 2
+    assert table2.get_string("b") == 2
     assert new_lookups.to_bytes() == lookups_bytes
 
 
@@ -72,10 +72,11 @@ def test_lookups_to_from_disk():
     assert "table2" in new_lookups
     table1 = new_lookups.get_table("table1")
     assert len(table1) == 2
-    assert table1.get("foo") == "bar"
+    assert table1.get_string("foo") == "bar"
     table2 = new_lookups.get_table("table2")
     assert len(table2) == 3
-    assert table2.get("b") == 2
+    assert table2.get_string("b") == 2
+
 
 
 # This fails on Python 3.5
@@ -93,10 +94,9 @@ def test_lookups_to_from_bytes_via_vocab():
     assert table_name in new_vocab.lookups
     table = new_vocab.lookups.get_table(table_name)
     assert len(table) == 2
-    assert table.get("hello") == "world"
+    assert table.get_string("hello") == "world"
     assert new_vocab.to_bytes() == vocab_bytes
 
-
 # This fails on Python 3.5
 @pytest.mark.xfail
 def test_lookups_to_from_disk_via_vocab():
@@ -113,4 +113,4 @@ def test_lookups_to_from_disk_via_vocab():
     assert table_name in new_vocab.lookups
     table = new_vocab.lookups.get_table(table_name)
     assert len(table) == 2
-    assert table.get("hello") == "world"
+    assert table.get_string("hello") == "world"

spacy/tokens/token.pyx

@@ -335,7 +335,7 @@ cdef class Token:
         """
         def __get__(self):
             if self.c.lemma == 0:
-                lemma_ = self.vocab.morphology.lemmatizer.lookup(self.orth_)
+                lemma_ = self.vocab.morphology.lemmatizer.lookup(self.orth, self.orth_)
                 return self.vocab.strings[lemma_]
             else:
                 return self.c.lemma
@@ -862,7 +862,7 @@ cdef class Token:
         """
         def __get__(self):
             if self.c.lemma == 0:
-                return self.vocab.morphology.lemmatizer.lookup(self.orth_)
+                return self.vocab.morphology.lemmatizer.lookup(self.orth, self.orth_)
             else:
                 return self.vocab.strings[self.c.lemma]
 

spacy/vocab.pyx

@@ -18,10 +18,10 @@ from .structs cimport SerializedLexemeC
 from .compat import copy_reg, basestring_
 from .errors import Errors
 from .lemmatizer import Lemmatizer
-from .lookups import Lookups
 from .attrs import intify_attrs, NORM
 from .vectors import Vectors
 from ._ml import link_vectors_to_models
+from .lookups import Lookups
 from . import util