* More docs work

spacy/lexeme.pxd

@@ -22,6 +22,7 @@ cdef struct Lexeme:
     StringHash* string_views
 
 
+cpdef StringHash lex_of(LexID lex_id) except 0
 cpdef char first_of(LexID lex_id) except 0
 cpdef size_t length_of(LexID lex_id) except 0
 cpdef double prob_of(LexID lex_id) except 0

spacy/lexeme.pyx

@@ -29,26 +29,21 @@ cpdef StringHash view_of(LexID lex_id, size_t view) except 0:
     return (<Lexeme*>lex_id).string_views[view]
 
 
-cpdef StringHash lex_of(size_t lex_id) except 0:
+cpdef StringHash lex_of(LexID lex_id) except 0:
-    '''Access the `lex' field of the Lexeme pointed to by lex_id.
+    '''Access a hash of the word's string.
 
-    The lex field is the hash of the string you would expect to get back from
+    >>> lex_of(lookup(u'Hi')) == hash(u'Hi')
-    a standard tokenizer, i.e. the word with punctuation and other non-whitespace
+    True
-    delimited tokens split off.  The other fields refer to properties of the
-    string that the lex field stores a hash of, except sic and tail.
-
-    >>> from spacy import en
-    >>> [en.unhash(lex_of(lex_id) for lex_id in en.tokenize(u'Hi! world')]
-    [u'Hi', u'!', u'world']
     '''
     return (<Lexeme*>lex_id).lex
 
 
 cpdef ClusterID cluster_of(LexID lex_id) except 0:
-    '''Access the `cluster' field of the Lexeme pointed to by lex_id, which
+    '''Access an integer representation of the word's Brown cluster.
-    gives an integer representation of the cluster ID of the word, 
-    which should be understood as a binary address:
 
+    A Brown cluster is an address into a binary tree, which gives some (noisy)
+    information about the word's distributional context.
+    
     >>> strings = (u'pineapple', u'apple', u'dapple', u'scalable')
     >>> token_ids = [lookup(s) for s in strings]
     >>> clusters = [cluster_of(t) for t in token_ids]
@@ -64,29 +59,28 @@ cpdef ClusterID cluster_of(LexID lex_id) except 0:
 
 
 cpdef char first_of(size_t lex_id) except 0:
-    '''Access the `first' field of the Lexeme pointed to by lex_id, which
+    '''Access the first byte of a utf8 encoding of the word.
-    stores the first character of the lex string of the word.
 
     >>> lex_id = lookup(u'Hello')
-    >>> unhash(first_of(lex_id))
+    >>> chr(first_of(lex_id))
-    u'H'
+    'H'
     '''
     return (<Lexeme*>lex_id).string[0]
 
 
 cpdef size_t length_of(size_t lex_id) except 0:
-    '''Access the `length' field of the Lexeme pointed to by lex_id, which stores
+    '''Access the (unicode) length of the word.
-    the length of the string hashed by lex_of.'''
+    '''
     cdef Lexeme* word = <Lexeme*>lex_id
     return word.length
 
 
 cpdef double prob_of(size_t lex_id) except 0:
-    '''Access the `prob' field of the Lexeme pointed to by lex_id, which stores
+    '''Access an estimate of the word's unigram log probability.
-    the smoothed unigram log probability of the word, as estimated from a large
+
-    text corpus.  By default, probabilities are based on counts from Gigaword,
+    Probabilities are calculated from a large text corpus, and smoothed using
-    smoothed using Knesser-Ney; but any probabilities file can be supplied to
+    simple Good-Turing.  Estimates are read from data/en/probabilities, and
-    load_probs.
+    can be replaced using spacy.en.load_probabilities.
     
     >>> prob_of(lookup(u'world'))
     -20.10340371976182
@@ -97,31 +91,39 @@ DEF OFT_UPPER = 1
 DEF OFT_TITLE = 2
 
 cpdef bint is_oft_upper(size_t lex_id):
-    '''Access the `oft_upper' field of the Lexeme pointed to by lex_id, which
+    '''Check the OFT_UPPER distributional flag for the word.
-    stores whether the lowered version of the string hashed by `lex' is found
+    
-    in all-upper case frequently in a large sample of text.  Users are free
+    The OFT_UPPER flag records whether a lower-cased version of the word
-    to load different data, by default we use a sample from Wikipedia, with
+    is found in all-upper case frequently in a large sample of text, where
-    a threshold of 0.95, picked to maximize mutual information for POS tagging.
+    "frequently" is defined as P >= 0.95 (chosen for high mutual information for
-
+    POS tagging).
-    >>> is_oft_upper(lookup(u'abc'))
+    
-    True
+    Case statistics are estimated from a large text corpus. Estimates are read
-    >>> is_oft_upper(lookup(u'aBc')) # This must get the same answer
+    from data/en/case_stats, and can be replaced using spacy.en.load_case_stats.
+    
+    >>> is_oft_upper(lookup(u'nato'))
     True
+    >>> is_oft_upper(lookup(u'the')) 
+    False
     '''
     return (<Lexeme*>lex_id).dist_flags & (1 << OFT_UPPER)
 
 
 cpdef bint is_oft_title(size_t lex_id):
-    '''Access the `oft_upper' field of the Lexeme pointed to by lex_id, which
+    '''Check the OFT_TITLE distributional flag for the word.
-    stores whether the lowered version of the string hashed by `lex' is found
+    
-    title-cased frequently in a large sample of text.  Users are free
+    The OFT_TITLE flag records whether a lower-cased version of the word
-    to load different data, by default we use a sample from Wikipedia, with
+    is found title-cased (see string.istitle) frequently in a large sample of text,
-    a threshold of 0.3, picked to maximize mutual information for POS tagging.
+    where "frequently" is defined as P >= 0.3 (chosen for high mutual information for
-
+    POS tagging).
-    >>> is_oft_title(lookup(u'marcus'))
+    
-    True
+    Case statistics are estimated from a large text corpus. Estimates are read
-    >>> is_oft_title(lookup(u'MARCUS')) # This must get the same value
+    from data/en/case_stats, and can be replaced using spacy.en.load_case_stats.
+    
+    >>> is_oft_upper(lookup(u'john'))
     True
+    >>> is_oft_upper(lookup(u'Bill')) 
+    False
     '''
     return (<Lexeme*>lex_id).dist_flags & (1 << OFT_TITLE)
 

spacy/orthography/latin.pyx

@@ -141,7 +141,11 @@ cpdef bint is_ascii(LexID lex_id) except *:
 
 
 cpdef StringHash norm_of(LexID lex_id) except 0:
-    """Return the hash of a normalized version of the string.
+    """Return the hash of a "normalized" version of the string.
+
+    Normalized strings are intended to be less sparse, while still capturing
+    important lexical information.  See spacy.latin.orthography.normalize_string for details of the normalization
+    function.
 
     >>> unhash(norm_of(lookupu'Hi'))
     u'hi'
@@ -154,7 +158,11 @@ cpdef StringHash norm_of(LexID lex_id) except 0:
 
 
 cpdef StringHash shape_of(LexID lex_id) except 0:
-    """Return the hash of the string shape.
+    """Return the hash of a string describing the word's "orthograpgic shape".
+
+    Orthographic shapes are calculated by the spacy.orthography.latin.string_shape
+    function. Word shape features have been found useful for NER and POS tagging,
+    e.g. Manning (2011)
 
     >>> unhash(shape_of(lookupu'Hi'))
     u'Xx'
@@ -168,8 +176,8 @@ cpdef StringHash shape_of(LexID lex_id) except 0:
 
 
 cpdef StringHash last3_of(LexID lex_id) except 0:
-    '''Access the `last3' field of the Lexeme pointed to by lex_id, which stores
+    '''Return the hash of string[-3:], i.e. the last three characters of the word.
-    the hash of the last three characters of the word:
+
     >>> lex_ids = [lookup(w) for w in (u'Hello', u'!')]
     >>> [unhash(last3_of(lex_id)) for lex_id in lex_ids]
     [u'llo', u'!']

spacy/util.py

@@ -1,6 +1,7 @@
 import os
 from os import path
 import codecs
+import json
 
 DATA_DIR = path.join(path.dirname(__file__), '..', 'data')
 
@@ -19,9 +20,13 @@ def load_case_stats(data_dir):
     return case_stats
 
 
-def load_dist_info(lang):
+def read_dist_info(lang):
-    with path.join(DATA_DIR, lang, 'distribution_info.json') as file_:
+    dist_path = path.join(DATA_DIR, lang, 'distribution_info.json')
-        dist_info = json.load(file_)
+    if path.exists(dist_path):
+        with open(dist_path) as file_:
+            dist_info = json.load(file_)
+    else:
+        dist_info = {}
     return dist_info