|
1
|
|
|
"""Backend that returns most similar subjects based on similarity in sparse |
|
2
|
|
|
TF-IDF normalized bag-of-words vector space""" |
|
3
|
|
|
|
|
4
|
|
|
import os.path |
|
5
|
|
|
import tempfile |
|
6
|
|
|
import joblib |
|
7
|
|
|
import gensim.similarities |
|
8
|
|
|
from gensim.matutils import Sparse2Corpus |
|
9
|
|
|
from sklearn.feature_extraction.text import TfidfVectorizer |
|
10
|
|
|
import annif.util |
|
11
|
|
|
from annif.suggestion import VectorSuggestionResult |
|
12
|
|
|
from annif.exception import NotInitializedException, NotSupportedException |
|
13
|
|
|
from . import backend |
|
14
|
|
|
|
|
15
|
|
|
|
|
16
|
|
|
class SubjectBuffer: |
|
17
|
|
|
"""A file-backed buffer to store and retrieve subject text.""" |
|
18
|
|
|
|
|
19
|
|
|
BUFFER_SIZE = 100 |
|
20
|
|
|
|
|
21
|
|
|
def __init__(self, tempdir, subject_id): |
|
22
|
|
|
filename = '{:08d}.txt'.format(subject_id) |
|
23
|
|
|
self._path = os.path.join(tempdir, filename) |
|
24
|
|
|
self._buffer = [] |
|
25
|
|
|
self._created = False |
|
26
|
|
|
|
|
27
|
|
|
def flush(self): |
|
28
|
|
|
if self._created: |
|
29
|
|
|
mode = 'a' |
|
30
|
|
|
else: |
|
31
|
|
|
mode = 'w' |
|
32
|
|
|
|
|
33
|
|
|
with open(self._path, mode, encoding='utf-8') as subjfile: |
|
34
|
|
|
for text in self._buffer: |
|
35
|
|
|
print(text, file=subjfile) |
|
36
|
|
|
|
|
37
|
|
|
self._buffer = [] |
|
38
|
|
|
self._created = True |
|
39
|
|
|
|
|
40
|
|
|
def write(self, text): |
|
41
|
|
|
self._buffer.append(text) |
|
42
|
|
|
if len(self._buffer) >= self.BUFFER_SIZE: |
|
43
|
|
|
self.flush() |
|
44
|
|
|
|
|
45
|
|
|
def read(self): |
|
46
|
|
|
if not self._created: |
|
47
|
|
|
# file was never created - we can simply return the buffer content |
|
48
|
|
|
return "\n".join(self._buffer) |
|
49
|
|
|
else: |
|
50
|
|
|
with open(self._path, 'r', encoding='utf-8') as subjfile: |
|
51
|
|
|
return subjfile.read() + "\n" + "\n".join(self._buffer) |
|
52
|
|
|
|
|
53
|
|
|
|
|
54
|
|
|
class TFIDFBackend(backend.AnnifBackend): |
|
55
|
|
|
"""TF-IDF vector space similarity based backend for Annif""" |
|
56
|
|
|
name = "tfidf" |
|
57
|
|
|
needs_subject_index = True |
|
58
|
|
|
|
|
59
|
|
|
# defaults for uninitialized instances |
|
60
|
|
|
_vectorizer = None |
|
61
|
|
|
_index = None |
|
62
|
|
|
|
|
63
|
|
|
VECTORIZER_FILE = 'vectorizer' |
|
64
|
|
|
INDEX_FILE = 'tfidf-index' |
|
65
|
|
|
|
|
66
|
|
|
def _generate_subjects_from_documents(self, corpus, project): |
|
67
|
|
|
with tempfile.TemporaryDirectory() as tempdir: |
|
68
|
|
|
subject_buffer = {} |
|
69
|
|
|
for subject_id in range(len(project.subjects)): |
|
70
|
|
|
subject_buffer[subject_id] = SubjectBuffer(tempdir, |
|
71
|
|
|
subject_id) |
|
72
|
|
|
|
|
73
|
|
|
for doc in corpus.documents: |
|
74
|
|
|
tokens = project.analyzer.tokenize_words(doc.text) |
|
75
|
|
|
for uri in doc.uris: |
|
76
|
|
|
subject_id = project.subjects.by_uri(uri) |
|
77
|
|
|
if subject_id is None: |
|
78
|
|
|
continue |
|
79
|
|
|
subject_buffer[subject_id].write(" ".join(tokens)) |
|
80
|
|
|
|
|
81
|
|
|
for sid in range(len(project.subjects)): |
|
82
|
|
|
yield subject_buffer[sid].read() |
|
83
|
|
|
|
|
84
|
|
|
def _initialize_vectorizer(self): |
|
85
|
|
|
if self._vectorizer is None: |
|
86
|
|
|
path = os.path.join(self.datadir, self.VECTORIZER_FILE) |
|
87
|
|
|
if os.path.exists(path): |
|
88
|
|
|
self.debug('loading vectorizer from {}'.format(path)) |
|
89
|
|
|
self._vectorizer = joblib.load(path) |
|
90
|
|
|
else: |
|
91
|
|
|
raise NotInitializedException( |
|
92
|
|
|
"vectorizer file '{}' not found".format(path), |
|
93
|
|
|
backend_id=self.backend_id) |
|
94
|
|
|
|
|
95
|
|
|
def _initialize_index(self): |
|
96
|
|
|
if self._index is None: |
|
97
|
|
|
path = os.path.join(self.datadir, self.INDEX_FILE) |
|
98
|
|
|
self.debug('loading similarity index from {}'.format(path)) |
|
99
|
|
|
if os.path.exists(path): |
|
100
|
|
|
self._index = gensim.similarities.SparseMatrixSimilarity.load( |
|
101
|
|
|
path) |
|
102
|
|
|
else: |
|
103
|
|
|
raise NotInitializedException( |
|
104
|
|
|
'similarity index {} not found'.format(path), |
|
105
|
|
|
backend_id=self.backend_id) |
|
106
|
|
|
|
|
107
|
|
|
def initialize(self): |
|
108
|
|
|
self._initialize_vectorizer() |
|
109
|
|
|
self._initialize_index() |
|
110
|
|
|
|
|
111
|
|
|
def _create_index(self, veccorpus): |
|
112
|
|
|
self.info('creating similarity index') |
|
113
|
|
|
gscorpus = Sparse2Corpus(veccorpus, documents_columns=False) |
|
114
|
|
|
self._index = gensim.similarities.SparseMatrixSimilarity( |
|
115
|
|
|
gscorpus, num_features=len(self._vectorizer.vocabulary_)) |
|
116
|
|
|
annif.util.atomic_save( |
|
117
|
|
|
self._index, |
|
118
|
|
|
self.datadir, |
|
119
|
|
|
self.INDEX_FILE) |
|
120
|
|
|
|
|
121
|
|
|
def train(self, corpus, project): |
|
122
|
|
|
if corpus.is_empty(): |
|
123
|
|
|
raise NotSupportedException( |
|
124
|
|
|
'Cannot train tfidf project with no documents') |
|
125
|
|
|
self.info('transforming subject corpus') |
|
126
|
|
|
subjects = self._generate_subjects_from_documents(corpus, project) |
|
127
|
|
|
self.info('creating vectorizer') |
|
128
|
|
|
self._vectorizer = TfidfVectorizer() |
|
129
|
|
|
veccorpus = self._vectorizer.fit_transform(subjects) |
|
130
|
|
|
annif.util.atomic_save( |
|
131
|
|
|
self._vectorizer, |
|
132
|
|
|
self.datadir, |
|
133
|
|
|
self.VECTORIZER_FILE, |
|
134
|
|
|
method=joblib.dump) |
|
135
|
|
|
self._create_index(veccorpus) |
|
136
|
|
|
|
|
137
|
|
|
def _suggest(self, text, project, params): |
|
138
|
|
|
self.debug('Suggesting subjects for text "{}..." (len={})'.format( |
|
139
|
|
|
text[:20], len(text))) |
|
140
|
|
|
tokens = project.analyzer.tokenize_words(text) |
|
141
|
|
|
vectors = self._vectorizer.transform([" ".join(tokens)]) |
|
142
|
|
|
docsim = self._index[vectors[0]] |
|
143
|
|
|
fullresult = VectorSuggestionResult(docsim, project.subjects) |
|
144
|
|
|
return fullresult.filter(limit=int(self.params['limit'])) |
|
145
|
|
|
|
NatLibFi /
Annif
