|
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 load_npz, save_npz |
|
11
|
|
|
from sklearn.preprocessing import normalize |
|
12
|
|
|
|
|
13
|
|
|
import annif.util |
|
14
|
|
|
from annif.exception import NotInitializedException, NotSupportedException |
|
15
|
|
|
from annif.suggestion import vector_to_suggestions |
|
16
|
|
|
|
|
17
|
|
|
from . import backend, mixins |
|
18
|
|
|
|
|
19
|
|
|
if TYPE_CHECKING: |
|
20
|
|
|
from collections.abc import Iterator |
|
21
|
|
|
|
|
22
|
|
|
from annif.corpus import Document, DocumentCorpus |
|
23
|
|
|
|
|
24
|
|
|
|
|
25
|
|
|
class SubjectBuffer: |
|
26
|
|
|
"""A file-backed buffer to store and retrieve subject text.""" |
|
27
|
|
|
|
|
28
|
|
|
BUFFER_SIZE = 100 |
|
29
|
|
|
|
|
30
|
|
|
def __init__(self, tempdir: str, subject_id: int) -> None: |
|
31
|
|
|
filename = "{:08d}.txt".format(subject_id) |
|
32
|
|
|
self._path = os.path.join(tempdir, filename) |
|
33
|
|
|
self._buffer = [] |
|
34
|
|
|
self._created = False |
|
35
|
|
|
|
|
36
|
|
|
def flush(self) -> None: |
|
37
|
|
|
if self._created: |
|
38
|
|
|
mode = "a" |
|
39
|
|
|
else: |
|
40
|
|
|
mode = "w" |
|
41
|
|
|
|
|
42
|
|
|
with open(self._path, mode, encoding="utf-8") as subjfile: |
|
43
|
|
|
for text in self._buffer: |
|
44
|
|
|
print(text, file=subjfile) |
|
45
|
|
|
|
|
46
|
|
|
self._buffer = [] |
|
47
|
|
|
self._created = True |
|
48
|
|
|
|
|
49
|
|
|
def write(self, text: str) -> None: |
|
50
|
|
|
self._buffer.append(text) |
|
51
|
|
|
if len(self._buffer) >= self.BUFFER_SIZE: |
|
52
|
|
|
self.flush() |
|
53
|
|
|
|
|
54
|
|
|
def read(self) -> str: |
|
55
|
|
|
if not self._created: |
|
56
|
|
|
# file was never created - we can simply return the buffer content |
|
57
|
|
|
return "\n".join(self._buffer) |
|
58
|
|
|
else: |
|
59
|
|
|
with open(self._path, "r", encoding="utf-8") as subjfile: |
|
60
|
|
|
return subjfile.read() + "\n" + "\n".join(self._buffer) |
|
61
|
|
|
|
|
62
|
|
|
|
|
63
|
|
|
class TFIDFBackend(mixins.TfidfVectorizerMixin, backend.AnnifBackend): |
|
64
|
|
|
"""TF-IDF vector space similarity based backend for Annif""" |
|
65
|
|
|
|
|
66
|
|
|
name = "tfidf" |
|
67
|
|
|
|
|
68
|
|
|
# defaults for uninitialized instances |
|
69
|
|
|
_tfidf_matrix = None |
|
70
|
|
|
|
|
71
|
|
|
MATRIX_FILE = "tfidf-matrix.npz" |
|
72
|
|
|
|
|
73
|
|
|
def _generate_subjects_from_documents( |
|
74
|
|
|
self, corpus: DocumentCorpus |
|
75
|
|
|
) -> Iterator[str]: |
|
76
|
|
|
with tempfile.TemporaryDirectory() as tempdir: |
|
77
|
|
|
subject_buffer = {} |
|
78
|
|
|
for subject_id in range(len(self.project.subjects)): |
|
79
|
|
|
subject_buffer[subject_id] = SubjectBuffer(tempdir, subject_id) |
|
80
|
|
|
|
|
81
|
|
|
for doc in corpus.documents: |
|
82
|
|
|
tokens = self.project.analyzer.tokenize_words(doc.text) |
|
83
|
|
|
for subject_id in doc.subject_set: |
|
84
|
|
|
subject_buffer[subject_id].write(" ".join(tokens)) |
|
85
|
|
|
|
|
86
|
|
|
for sid in range(len(self.project.subjects)): |
|
87
|
|
|
yield subject_buffer[sid].read() |
|
88
|
|
|
|
|
89
|
|
|
def _initialize_index(self) -> None: |
|
90
|
|
|
if self._tfidf_matrix is None: |
|
91
|
|
|
path = os.path.join(self.datadir, self.MATRIX_FILE) |
|
92
|
|
|
self.debug("loading tf-idf matrix from {}".format(path)) |
|
93
|
|
|
if os.path.exists(path): |
|
94
|
|
|
self._tfidf_matrix = load_npz(path) |
|
95
|
|
|
else: |
|
96
|
|
|
raise NotInitializedException( |
|
97
|
|
|
"tf-idf matrix {} not found".format(path), |
|
98
|
|
|
backend_id=self.backend_id, |
|
99
|
|
|
) |
|
100
|
|
|
|
|
101
|
|
|
def initialize(self, parallel: bool = False) -> None: |
|
102
|
|
|
self.initialize_vectorizer() |
|
103
|
|
|
self._initialize_index() |
|
104
|
|
|
|
|
105
|
|
|
def _train( |
|
106
|
|
|
self, |
|
107
|
|
|
corpus: DocumentCorpus, |
|
108
|
|
|
params: dict[str, Any], |
|
109
|
|
|
jobs: int = 0, |
|
110
|
|
|
) -> None: |
|
111
|
|
|
if corpus == "cached": |
|
112
|
|
|
raise NotSupportedException( |
|
113
|
|
|
"Training tfidf project from cached data not supported." |
|
114
|
|
|
) |
|
115
|
|
|
if corpus.is_empty(): |
|
116
|
|
|
raise NotSupportedException("Cannot train tfidf project with no documents") |
|
117
|
|
|
self.info("transforming subject corpus") |
|
118
|
|
|
subjects = self._generate_subjects_from_documents(corpus) |
|
119
|
|
|
self._tfidf_matrix = normalize(self.create_vectorizer(subjects)) |
|
120
|
|
|
self.info("saving tf-idf matrix") |
|
121
|
|
|
annif.util.atomic_save( |
|
122
|
|
|
self._tfidf_matrix, |
|
123
|
|
|
self.datadir, |
|
124
|
|
|
self.MATRIX_FILE, |
|
125
|
|
|
lambda obj, filename: save_npz(filename, obj), |
|
126
|
|
|
) |
|
127
|
|
|
|
|
128
|
|
|
def _suggest(self, doc: Document, params: dict[str, Any]) -> Iterator: |
|
129
|
|
|
self.debug( |
|
130
|
|
|
'Suggesting subjects for text "{}..." (len={})'.format( |
|
131
|
|
|
doc.text[:20], len(doc.text) |
|
132
|
|
|
) |
|
133
|
|
|
) |
|
134
|
|
|
tokens = self.project.analyzer.tokenize_words(doc.text) |
|
135
|
|
|
query_vector = normalize(self.vectorizer.transform([" ".join(tokens)])) |
|
136
|
|
|
|
|
137
|
|
|
# Compute cosine similarity between query and indexed corpus |
|
138
|
|
|
similarities = (query_vector @ self._tfidf_matrix.T).toarray().flatten() |
|
139
|
|
|
|
|
140
|
|
|
return vector_to_suggestions(similarities, int(params["limit"])) |
|
141
|
|
|
|
NatLibFi /
Annif
