1
|
|
|
"""Backend that returns most similar subjects based on similarity in sparse |
2
|
|
|
TF-IDF normalized bag-of-words vector space""" |
3
|
|
|
|
4
|
|
|
from __future__ import annotations |
5
|
|
|
|
6
|
|
|
import os.path |
7
|
|
|
import tempfile |
8
|
|
|
from typing import TYPE_CHECKING, Any |
9
|
|
|
|
10
|
|
|
from scipy.sparse import csr_array, load_npz, save_npz |
11
|
|
|
from sklearn.preprocessing import normalize |
12
|
|
|
|
13
|
|
|
import annif.util |
14
|
|
|
from annif.exception import ( |
15
|
|
|
NotInitializedException, |
16
|
|
|
NotSupportedException, |
17
|
|
|
OperationFailedException, |
18
|
|
|
) |
19
|
|
|
from annif.suggestion import SuggestionBatch |
20
|
|
|
|
21
|
|
|
from . import backend, mixins |
22
|
|
|
|
23
|
|
|
if TYPE_CHECKING: |
24
|
|
|
from collections.abc import Iterator |
25
|
|
|
|
26
|
|
|
from annif.corpus import Document, DocumentCorpus |
27
|
|
|
|
28
|
|
|
|
29
|
|
|
class SubjectBuffer: |
30
|
|
|
"""A file-backed buffer to store and retrieve subject text.""" |
31
|
|
|
|
32
|
|
|
BUFFER_SIZE = 100 |
33
|
|
|
|
34
|
|
|
def __init__(self, tempdir: str, subject_id: int) -> None: |
35
|
|
|
filename = "{:08d}.txt".format(subject_id) |
36
|
|
|
self._path = os.path.join(tempdir, filename) |
37
|
|
|
self._buffer = [] |
38
|
|
|
self._created = False |
39
|
|
|
|
40
|
|
|
def flush(self) -> None: |
41
|
|
|
if self._created: |
42
|
|
|
mode = "a" |
43
|
|
|
else: |
44
|
|
|
mode = "w" |
45
|
|
|
|
46
|
|
|
with open(self._path, mode, encoding="utf-8") as subjfile: |
47
|
|
|
for text in self._buffer: |
48
|
|
|
print(text, file=subjfile) |
49
|
|
|
|
50
|
|
|
self._buffer = [] |
51
|
|
|
self._created = True |
52
|
|
|
|
53
|
|
|
def write(self, text: str) -> None: |
54
|
|
|
self._buffer.append(text) |
55
|
|
|
if len(self._buffer) >= self.BUFFER_SIZE: |
56
|
|
|
self.flush() |
57
|
|
|
|
58
|
|
|
def read(self) -> str: |
59
|
|
|
if not self._created: |
60
|
|
|
# file was never created - we can simply return the buffer content |
61
|
|
|
return "\n".join(self._buffer) |
62
|
|
|
else: |
63
|
|
|
with open(self._path, "r", encoding="utf-8") as subjfile: |
64
|
|
|
return subjfile.read() + "\n" + "\n".join(self._buffer) |
65
|
|
|
|
66
|
|
|
|
67
|
|
|
class TFIDFBackend(mixins.TfidfVectorizerMixin, backend.AnnifBackend): |
68
|
|
|
"""TF-IDF vector space similarity based backend for Annif""" |
69
|
|
|
|
70
|
|
|
name = "tfidf" |
71
|
|
|
|
72
|
|
|
# defaults for uninitialized instances |
73
|
|
|
_tfidf_matrix = None |
74
|
|
|
|
75
|
|
|
MATRIX_FILE = "tfidf-matrix.npz" |
76
|
|
|
OLD_INDEX_FILE = "tfidf-index" |
77
|
|
|
|
78
|
|
|
def _generate_subjects_from_documents( |
79
|
|
|
self, corpus: DocumentCorpus |
80
|
|
|
) -> Iterator[str]: |
81
|
|
|
with tempfile.TemporaryDirectory() as tempdir: |
82
|
|
|
subject_buffer = {} |
83
|
|
|
for subject_id in range(len(self.project.subjects)): |
84
|
|
|
subject_buffer[subject_id] = SubjectBuffer(tempdir, subject_id) |
85
|
|
|
|
86
|
|
|
for doc in corpus.documents: |
87
|
|
|
tokens = self.project.analyzer.tokenize_words(doc.text) |
88
|
|
|
for subject_id in doc.subject_set: |
89
|
|
|
subject_buffer[subject_id].write(" ".join(tokens)) |
90
|
|
|
|
91
|
|
|
for sid in range(len(self.project.subjects)): |
92
|
|
|
yield subject_buffer[sid].read() |
93
|
|
|
|
94
|
|
|
def _initialize_index(self) -> None: |
95
|
|
|
if self._tfidf_matrix is None: |
96
|
|
|
path = os.path.join(self.datadir, self.MATRIX_FILE) |
97
|
|
|
self.debug("loading tf-idf matrix from {}".format(path)) |
98
|
|
|
if os.path.exists(path): |
99
|
|
|
self._tfidf_matrix = load_npz(path) |
100
|
|
|
elif os.path.exists(os.path.join(self.datadir, self.OLD_INDEX_FILE)): |
101
|
|
|
raise OperationFailedException( |
102
|
|
|
"TFIDF models trained on Annif versions older than 1.4 cannot be " |
103
|
|
|
"loaded. Please retrain your project." |
104
|
|
|
) |
105
|
|
|
else: |
106
|
|
|
raise NotInitializedException( |
107
|
|
|
"tf-idf matrix {} not found".format(path), |
108
|
|
|
backend_id=self.backend_id, |
109
|
|
|
) |
110
|
|
|
|
111
|
|
|
def initialize(self, parallel: bool = False) -> None: |
112
|
|
|
self.initialize_vectorizer() |
113
|
|
|
self._initialize_index() |
114
|
|
|
|
115
|
|
|
def _train( |
116
|
|
|
self, |
117
|
|
|
corpus: DocumentCorpus, |
118
|
|
|
params: dict[str, Any], |
119
|
|
|
jobs: int = 0, |
120
|
|
|
) -> None: |
121
|
|
|
if corpus == "cached": |
122
|
|
|
raise NotSupportedException( |
123
|
|
|
"Training tfidf project from cached data not supported." |
124
|
|
|
) |
125
|
|
|
if corpus.is_empty(): |
126
|
|
|
raise NotSupportedException("Cannot train tfidf project with no documents") |
127
|
|
|
self.info("transforming subject corpus") |
128
|
|
|
subjects = self._generate_subjects_from_documents(corpus) |
129
|
|
|
# Note: Intentionally don't pass a tokenizer to the vectorizer here. |
130
|
|
|
# Instead the tokenization is done inside _generate_subjects_from_documents |
131
|
|
|
# and in _suggest_batch. This way, the same train document doesn't have to be |
132
|
|
|
# tokenized many times during training if it has many subjects. |
133
|
|
|
self._tfidf_matrix = normalize(self.create_vectorizer(subjects)) |
134
|
|
|
self.info("saving tf-idf matrix") |
135
|
|
|
annif.util.atomic_save( |
136
|
|
|
self._tfidf_matrix, |
137
|
|
|
self.datadir, |
138
|
|
|
self.MATRIX_FILE, |
139
|
|
|
lambda obj, filename: save_npz(filename, obj), |
140
|
|
|
) |
141
|
|
|
|
142
|
|
|
def _suggest_batch( |
143
|
|
|
self, documents: list[Document], params: dict[str, Any] |
144
|
|
|
) -> SuggestionBatch: |
145
|
|
|
query_vector = normalize( |
146
|
|
|
self.vectorizer.transform( |
147
|
|
|
[ |
148
|
|
|
" ".join(self.project.analyzer.tokenize_words(doc.text)) |
149
|
|
|
for doc in documents |
150
|
|
|
] |
151
|
|
|
) |
152
|
|
|
) |
153
|
|
|
|
154
|
|
|
# Compute cosine similarity between query and indexed corpus |
155
|
|
|
similarities = query_vector @ self._tfidf_matrix.T |
156
|
|
|
|
157
|
|
|
return SuggestionBatch(csr_array(similarities)).filter(int(params["limit"])) |
158
|
|
|
|