File size: 7,657 Bytes
e4b9a7b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
# -*- coding: utf-8 -*-
# SPDX-FileCopyrightText: 2016-2025 PyThaiNLP Project
# SPDX-FileType: SOURCE
# SPDX-License-Identifier: Apache-2.0

import re
import sys
from typing import List, Tuple

import numpy as np
import pandas as pd

SEPARATOR = "|"

# regex for removing one space surrounded by separators, i.e. | |
SURROUNDING_SEPS_RX = re.compile(
    "{sep}? ?{sep}$".format(sep=re.escape(SEPARATOR))
)

# regex for removing repeated separators, i.e. ||||
MULTIPLE_SEPS_RX = re.compile("{sep}+".format(sep=re.escape(SEPARATOR)))

# regex for removing tags, i.e. <NE>, </NE>
TAG_RX = re.compile(r"<\/?[A-Z]+>")

# regex for removing trailing separators, i.e.  a|dog| -> a|dog
TAILING_SEP_RX = re.compile("{sep}$".format(sep=re.escape(SEPARATOR)))


def _f1(precision: float, recall: float) -> float:
    """
    Compute f1.

    :param float precision
    :param float recall

    :return: f1
    :rtype: float
    """
    if precision == recall == 0:
        return 0
    return 2 * precision * recall / (precision + recall)


def _flatten_result(my_dict: dict, sep: str = ":") -> dict:
    """
    Flatten two-dimension dictionary.

    Use keys in the first dimension as a prefix for keys in the second dimension.
    For example,
    my_dict = { "a": { "b": 7 } }
    flatten(my_dict)
    { "a:b": 7 }


    :param dict my_dict: dictionary containing stats
    :param str sep: separator between the two keys (default: ":")

    :return: a one-dimension dictionary with keys combined
    :rtype: dict[str, float | str]
    """
    items = []
    for k1, kv2 in my_dict.items():
        for k2, v in kv2.items():
            new_key = f"{k1}{sep}{k2}"
            items.append((new_key, v))

    return dict(items)


def benchmark(ref_samples: List[str], samples: List[str]) -> pd.DataFrame:
    """
    Performance benchmarking for samples.

    Please see :meth:`pythainlp.benchmarks.word_tokenization.compute_stats` for
    the computed metrics.

    :param list[str] ref_samples: ground truth for samples
    :param list[str] samples: samples that we want to evaluate

    :return: dataframe with row x col = len(samples) x len(metrics)
    :rtype: pandas.DataFrame
    """
    results = []
    for i, (r, s) in enumerate(zip(ref_samples, samples)):
        try:
            r, s = preprocessing(r), preprocessing(s)
            if r and s:
                stats = compute_stats(r, s)
                stats = _flatten_result(stats)
                stats["expected"] = r
                stats["actual"] = s
                results.append(stats)
        except:
            reason = """
[Error]
Reason: %s

Pair (i=%d)
--- label
%s
--- sample
%s
""" % (
                sys.exc_info(),
                i,
                r,
                s,
            )
            raise SystemExit(reason)

    return pd.DataFrame(results)


def preprocessing(txt: str, remove_space: bool = True) -> str:
    """
    Clean up text before performing evaluation.

    :param str text: text to be preprocessed
    :param bool remove_space: whether to remove white space

    :return: preprocessed text
    :rtype: str
    """
    txt = re.sub(SURROUNDING_SEPS_RX, "", txt)

    if remove_space:
        txt = re.sub(r"\s+", "", txt)

    txt = re.sub(MULTIPLE_SEPS_RX, SEPARATOR, txt)

    txt = re.sub(TAG_RX, "", txt)

    txt = re.sub(TAILING_SEP_RX, "", txt).strip()

    return txt


def compute_stats(ref_sample: str, raw_sample: str) -> dict:
    """
    Compute statistics for tokenization quality

    These statistics include:

    **Character-Level**:
      True Positive, False Positive, True Negative, False Negative, Precision, Recall, and f1
    **Word-Level**:
      Precision, Recall, and f1
    **Other**:
      - Correct tokenization indicator: {0, 1} sequence indicating that the corresponding
        word is tokenized correctly.

    :param str ref_sample: ground truth for samples
    :param str samples: samples that we want to evaluate

    :return: metrics at character- and word-level and indicators of correctly tokenized words
    :rtype: dict[str, float | str]
    """
    ref_sample = _binary_representation(ref_sample)
    sample = _binary_representation(raw_sample)

    # Compute character-level statistics
    c_pos_pred, c_neg_pred = np.argwhere(sample == 1), np.argwhere(sample == 0)

    c_pos_pred = c_pos_pred[c_pos_pred < ref_sample.shape[0]]
    c_neg_pred = c_neg_pred[c_neg_pred < ref_sample.shape[0]]

    c_tp = np.sum(ref_sample[c_pos_pred] == 1)
    c_fp = np.sum(ref_sample[c_pos_pred] == 0)

    c_tn = np.sum(ref_sample[c_neg_pred] == 0)
    c_fn = np.sum(ref_sample[c_neg_pred] == 1)

    # Compute word-level statistics

    # Find correctly tokenized words in the reference sample
    word_boundaries = _find_word_boundaries(ref_sample)

    # Find correctly tokenized words in the sample
    ss_boundaries = _find_word_boundaries(sample)
    tokenization_indicators = _find_words_correctly_tokenised(
        word_boundaries, ss_boundaries
    )

    correctly_tokenised_words = np.sum(tokenization_indicators)

    tokenization_indicators = list(
        map(str, tokenization_indicators)
    )

    return {
        "char_level": {
            "tp": c_tp,
            "fp": c_fp,
            "tn": c_tn,
            "fn": c_fn,
        },
        "word_level": {
            "correctly_tokenised_words": correctly_tokenised_words,
            "total_words_in_sample": np.sum(sample),
            "total_words_in_ref_sample": np.sum(ref_sample),
        },
        "global": {
            "tokenisation_indicators": "".join(tokenization_indicators)
        },
    }


def _binary_representation(txt: str, verbose: bool = False):
    """
    Transform text into {0, 1} sequence.

    where (1) indicates that the corresponding character is the beginning of
    a word. For example, ผม|ไม่|ชอบ|กิน|ผัก -> 10100...

    :param str txt: input text that we want to transform
    :param bool verbose: for debugging purposes

    :return: {0, 1} sequence
    :rtype: str
    """
    chars = np.array(list(txt))

    boundary = np.argwhere(chars == SEPARATOR).reshape(-1)
    boundary = boundary - np.array(range(boundary.shape[0]))

    bin_rept = np.zeros(len(txt) - boundary.shape[0])
    bin_rept[list(boundary) + [0]] = 1

    sample_wo_seps = list(txt.replace(SEPARATOR, ""))

    # sanity check
    assert len(sample_wo_seps) == len(bin_rept)

    if verbose:
        for c, m in zip(sample_wo_seps, bin_rept):
            print("%s -- %d" % (c, m))

    return bin_rept


def _find_word_boundaries(bin_reps) -> list:
    """
    Find the starting and ending location of each word.

    :param str bin_reps: binary representation of a text

    :return: list of tuples (start, end)
    :rtype: list[tuple(int, int)]
    """
    boundary = np.argwhere(bin_reps == 1).reshape(-1)
    start_idx = boundary
    end_idx = boundary[1:].tolist() + [bin_reps.shape[0]]

    return list(zip(start_idx, end_idx))


def _find_words_correctly_tokenised(
    ref_boundaries: List[Tuple[int, int]],
    predicted_boundaries: List[Tuple[int, int]],
) -> Tuple[int]:
    """
    Find whether each word is correctly tokenized.

    :param list[tuple(int, int)] ref_boundaries: word boundaries of reference tokenization
    :param list[tuple(int, int)] predicted_boundaries: word boundareies of predicted tokenization

    :return: binary sequence where 1 indicates the corresponding word is tokenized correctly
    :rtype: tuple[int]
    """
    ref_b = dict(zip(ref_boundaries, [1] * len(ref_boundaries)))

    labels = tuple(map(lambda x: ref_b.get(x, 0), predicted_boundaries))
    return labels