File size: 18,850 Bytes
f036197
 
 
 
 
 
 
 
 
 
 
 
 
64ec1e5
f583bc0
f036197
a249916
 
90d2f63
 
 
61ff8d5
90d2f63
 
 
de5dcb7
a249916
 
 
f036197
 
90d2f63
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
f036197
 
 
 
a1a6383
90d2f63
 
f036197
 
 
a1a6383
90d2f63
f036197
64ec1e5
 
f036197
90d2f63
 
64ec1e5
90d2f63
 
64ec1e5
 
 
 
 
 
 
f036197
dfd7508
 
 
 
64ec1e5
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
f036197
 
90d2f63
 
 
 
 
 
 
 
 
 
 
 
 
f036197
 
de5dcb7
a249916
 
 
90d2f63
a249916
 
 
90d2f63
a249916
 
 
 
 
 
 
90d2f63
a249916
 
90d2f63
a249916
 
90d2f63
a249916
 
 
90d2f63
a249916
90d2f63
 
a249916
 
 
 
 
 
 
 
90d2f63
a249916
 
 
 
 
90d2f63
a249916
 
90d2f63
a249916
 
 
 
 
f036197
a249916
 
 
f036197
a249916
 
 
de5dcb7
a249916
 
 
de5dcb7
a249916
 
de5dcb7
a249916
 
 
de5dcb7
a249916
de5dcb7
 
 
 
 
 
 
 
 
 
 
 
 
f036197
 
90d2f63
 
 
a249916
90d2f63
f036197
 
 
 
 
 
 
 
 
61ff8d5
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
90d2f63
f036197
90d2f63
f036197
 
90d2f63
4c522b3
a249916
4c522b3
90d2f63
 
 
 
 
 
 
61ff8d5
 
 
f6ab8c5
de5dcb7
 
f6ab8c5
61ff8d5
 
 
 
 
 
de5dcb7
a249916
61ff8d5
a249916
de5dcb7
 
 
 
 
 
 
a249916
 
 
 
de5dcb7
 
 
 
a249916
de5dcb7
 
 
 
 
 
 
 
a249916
 
 
61ff8d5
 
90d2f63
a249916
de5dcb7
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
90d2f63
61ff8d5
de5dcb7
90d2f63
de5dcb7
 
 
 
 
 
 
 
 
 
 
 
 
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
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
# Copyright 2020 The HuggingFace Datasets Authors and the current dataset script contributor.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# TODO: Add test cases, Remove tokenize_sentences flag since it can be determined from the input itself.
"""Sem-F1 metric"""

from functools import partial
from typing import List, Optional, Tuple

import datasets
import evaluate
import nltk
import numpy as np
from numpy.typing import NDArray
from sklearn.metrics.pairwise import cosine_similarity

from encoder_models import get_encoder
from type_aliases import DEVICE_TYPE, PREDICTION_TYPE, REFERENCE_TYPE
from utils import is_nested_list_of_type, Scores, slice_embeddings, flatten_list, get_gpu

_CITATION = """\
@inproceedings{bansal-etal-2022-sem,
    title = "{SEM}-F1: an Automatic Way for Semantic Evaluation of Multi-Narrative Overlap Summaries at Scale",
    author = "Bansal, Naman  and
      Akter, Mousumi  and
      Karmaker Santu, Shubhra Kanti",
    editor = "Goldberg, Yoav  and
      Kozareva, Zornitsa  and
      Zhang, Yue",
    booktitle = "Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing",
    month = dec,
    year = "2022",
    address = "Abu Dhabi, United Arab Emirates",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2022.emnlp-main.49",
    doi = "10.18653/v1/2022.emnlp-main.49",
    pages = "780--792",
    abstract = "Recent work has introduced an important yet relatively under-explored NLP task called Semantic Overlap Summarization (SOS) that entails generating a summary from multiple alternative narratives which conveys the common information provided by those narratives. Previous work also published a benchmark dataset for this task by collecting 2,925 alternative narrative pairs from the web and manually annotating 411 different reference summaries by engaging human annotators. In this paper, we exclusively focus on the automated evaluation of the SOS task using the benchmark dataset. More specifically, we first use the popular ROUGE metric from text-summarization literature and conduct a systematic study to evaluate the SOS task. Our experiments discover that ROUGE is not suitable for this novel task and therefore, we propose a new sentence-level precision-recall style automated evaluation metric, called SEM-F1 (Semantic F1). It is inspired by the benefits of the sentence-wise annotation technique using overlap labels reported by the previous work. Our experiments show that the proposed SEM-F1 metric yields a higher correlation with human judgment and higher inter-rater agreement compared to the ROUGE metric.",
}
"""

_DESCRIPTION = """\
Sem-F1 metric leverages the pre-trained contextual embeddings and evaluates the model generated 
semantic overlap summary with the reference overlap summary. It evaluates the semantic overlap summary at the 
sentence level and computes precision, recall and F1 scores.
"""

_KWARGS_DESCRIPTION = """
Sem-F1 compares the system generated overlap summary with ground truth reference overlap.

Args:
    predictions: list - List of  predictions (Details below)
    references: list - List of references (Details below)
        reference should be a string with tokens separated by spaces.
    model_type: str - Model to use. [pv1, stsb, use] 
        Options: 
            pv1 - paraphrase-distilroberta-base-v1 (Default)
            stsb - stsb-roberta-large 
            use - Universal Sentence Encoder
    tokenize_sentences: bool - Sentence tokenize the input document (prediction/reference). Default: True.
    gpu: Union[bool, int] - Whether to use GPU or CPU. 
        Options: 
            False - CPU (Default)
            True - GPU, device 0
            n: int - GPU, device n
    batch_size: int - Batch Size, Default = 32.
Returns:
    precision: Precision.
    recall: Recall.
    f1: F1 score.
    
There are 4 possible cases for inputs corresponding to predictions and references arguments 
Case 1: Multi-Ref = False, tokenize_sentences = False
    predictions: List[List[str]] - List of  predictions where each prediction is a list of sentences.
    references: List[List[str]] - List of references where each reference is a list of sentences.
Case 2: Multi-Ref = False, tokenize_sentences = True
    predictions: List[str] - List of  predictions where each prediction is a document
    references: List[str] - List of references where each reference is a document
Case 3: Multi-Ref = True, tokenize_sentences = False
    predictions: List[List[str]] - List of  predictions where each prediction is a list of sentences.
    references: List[List[List[str]]] - List of multi-references i.e. [[r11, r12, ...], [r21, r22, ...], ...] 
                                        where each rij is further a list of sentences
Case 4: Multi-Ref = True, tokenize_sentences = True
    predictions: List[str] - List of  predictions where each prediction is a document
    references: List[List[str]] - List of multi-references i.e. [[r11, r12, ...], [r21, r22, ...], ...]
                                  where each rij is a document  
    
This can be seen in the form of truth table as follows:
Case | Multi-Ref | tokenize_sentences | predictions     | references
1    | 0         | 0                  | List[List[str]] | List[List[str]]        
2    | 0         | 1                  | List[str]       | List[str]              
3    | 1         | 0                  | List[List[str]] | List[List[List[str]]]
4    | 1         | 1                  | List[str]       | List[List[str]]      

It is automatically determined whether it is Multi-Ref case Single-Ref case.
 
Examples:

    >>> import evaluate
    >>> predictions = [
        ["I go to School.", "You are stupid."],
        ["I love adventure sports."],
    ]
    >>> references = [
        ["I go to School.", "You are stupid."],
        ["I love adventure sports."],
    ]
    >>> metric = evaluate.load("semf1")
    >>> results = metric.compute(predictions=predictions, references=references)
    >>> print([round(v, 2) for v in results["f1"]])
    [0.77, 0.56]
"""


def _compute_cosine_similarity(pred_embeds: NDArray, ref_embeds: NDArray) -> Tuple[float, float]:
    """
        Compute precision and recall based on cosine similarity between predicted and reference embeddings.

        Args:
            pred_embeds (NDArray): Predicted embeddings (shape: [num_pred, embedding_dim]).
            ref_embeds (NDArray): Reference embeddings (shape: [num_ref, embedding_dim]).

        Returns:
            Tuple[float, float]: Precision and recall based on cosine similarity scores.
                Precision: Average maximum cosine similarity score per predicted embedding.
                Recall: Average maximum cosine similarity score per reference embedding.
        """
    # Compute cosine similarity between predicted and reference embeddings
    cosine_scores = cosine_similarity(pred_embeds, ref_embeds)

    # Compute precision per predicted embedding
    precision_per_sentence_sim = np.max(cosine_scores, axis=-1)

    # Compute recall per reference embedding
    recall_per_sentence_sim = np.max(cosine_scores, axis=0)

    # Calculate mean precision and recall scores
    precision = np.mean(precision_per_sentence_sim).item()
    recall = np.mean(recall_per_sentence_sim).item()

    return precision, recall


def _validate_input_format(
        tokenize_sentences: bool,
        multi_references: bool,
        predictions: PREDICTION_TYPE,
        references: REFERENCE_TYPE,
):
    """
        Validate the format of predictions and references based on specified criteria.

        Args:
        - tokenize_sentences (bool): Flag indicating whether sentences should be tokenized.
        - multi_references (bool): Flag indicating whether multiple references are provided.
        - predictions (PREDICTION_TYPE): Predictions to validate.
        - references (REFERENCE_TYPE): References to validate.

        Raises:
        - ValueError: If the format of predictions or references does not meet the specified criteria.

        Validation Criteria:
        The function validates predictions and references based on the following conditions:
        1. If `tokenize_sentences` is True and `multi_references` is True:
           - Predictions must be a list of strings (`is_list_of_strings_at_depth(predictions, 1)`).
           - References must be a list of list of strings (`is_list_of_strings_at_depth(references, 2)`).

        2. If `tokenize_sentences` is False and `multi_references` is True:
           - Predictions must be a list of list of strings (`is_list_of_strings_at_depth(predictions, 2)`).
           - References must be a list of list of list of strings (`is_list_of_strings_at_depth(references, 3)`).

        3. If `tokenize_sentences` is True and `multi_references` is False:
           - Predictions must be a list of strings (`is_list_of_strings_at_depth(predictions, 1)`).
           - References must be a list of strings (`is_list_of_strings_at_depth(references, 1)`).

        4. If `tokenize_sentences` is False and `multi_references` is False:
           - Predictions must be a list of list of strings (`is_list_of_strings_at_depth(predictions, 2)`).
           - References must be a list of list of strings (`is_list_of_strings_at_depth(references, 2)`).

        The function checks these conditions and raises a ValueError if any condition is not met,
        indicating that predictions or references are not in the valid input format.

        Note:
        - `PREDICTION_TYPE` and `REFERENCE_TYPE` are defined at the top of the file
    """

    is_list_of_strings_at_depth = partial(is_nested_list_of_type, element_type=str)
    if tokenize_sentences and multi_references:
        condition = is_list_of_strings_at_depth(predictions, 1) and is_list_of_strings_at_depth(references, 2)
    elif not tokenize_sentences and multi_references:
        condition = is_list_of_strings_at_depth(predictions, 2) and is_list_of_strings_at_depth(references, 3)
    elif tokenize_sentences and not multi_references:
        condition = is_list_of_strings_at_depth(predictions, 1) and is_list_of_strings_at_depth(references, 1)
    else:
        condition = is_list_of_strings_at_depth(predictions, 2) and is_list_of_strings_at_depth(references, 2)

    if not condition:
        raise ValueError("Predictions are references are not valid input format. Refer to documentation.")


@evaluate.utils.file_utils.add_start_docstrings(_DESCRIPTION, _KWARGS_DESCRIPTION)
class SemF1(evaluate.Metric):
    _MODEL_TYPE_TO_NAME = {
        "pv1": "paraphrase-distilroberta-base-v1",
        "stsb": "stsb-roberta-large",
        "use": "use",  # "sentence-transformers/use-cmlm-multilingual",  # TODO: check PyTorch USE VS TF USE
    }

    def _info(self):
        return evaluate.MetricInfo(
            # This is the description that will appear on the modules page.
            module_type="metric",
            description=_DESCRIPTION,
            citation=_CITATION,
            inputs_description=_KWARGS_DESCRIPTION,
            # This defines the format of each prediction and reference
            features=[
                # Multi References: False, Tokenize_Sentences = False
                datasets.Features(
                    {
                        # predictions: List[List[str]] - List of predictions where prediction is a list of sentences
                        "predictions": datasets.Sequence(datasets.Value("string", id="sequence"), id="predictions"),
                        # references: List[List[str]] - List of references where each reference is a list of sentences
                        "references": datasets.Sequence(datasets.Value("string", id="sequence"), id="references"),
                    }
                ),
                # Multi References: False, Tokenize_Sentences = True
                datasets.Features(
                    {
                        # predictions: List[str] - List of predictions
                        "predictions": datasets.Value("string", id="sequence"),
                        # references: List[str] - List of documents
                        "references": datasets.Value("string", id="sequence"),
                    }
                ),
                # Multi References: True, Tokenize_Sentences = False
                datasets.Features(
                    {
                        # predictions: List[List[str]] - List of predictions where prediction is a list of sentences
                        "predictions": datasets.Sequence(datasets.Value("string", id="sequence"), id="predictions"),
                        # references: List[List[List[str]]] - List of multi-references.
                        #                                     So each "reference" is also a list (r1, r2, ...).
                        #                                     Further, each ri's are also list of sentences.
                        "references": datasets.Sequence(
                            datasets.Sequence(datasets.Value("string", id="sequence"), id="ref"), id="references"),
                    }
                ),
                # Multi References: True, Tokenize_Sentences = True
                datasets.Features(
                    {
                        # predictions: List[str] - List of predictions
                        "predictions": datasets.Value("string", id="sequence"),
                        # references: List[List[List[str]]] - List of multi-references.
                        #                                     So each "reference" is also a list (r1, r2, ...).
                        "references": datasets.Sequence(datasets.Value("string", id="ref"), id="references"),
                    }
                ),
            ],
            # # Homepage of the module for documentation
            # Additional links to the codebase or references
            reference_urls=["https://aclanthology.org/2022.emnlp-main.49/"]
        )

    def _get_model_name(self, model_type: Optional[str] = None) -> str:
        if model_type is None:
            model_type = "use"

        if model_type not in self._MODEL_TYPE_TO_NAME.keys():
            raise ValueError(f"Provide a correct model_type.\n"
                             f"Options: {self._MODEL_TYPE_TO_NAME.keys()}\n"
                             f"Currently provided: {model_type}")

        return self._MODEL_TYPE_TO_NAME[model_type]

    def _download_and_prepare(self, dl_manager):
        """Optional: download external resources useful to compute the scores"""
        import nltk
        nltk.download("punkt", quiet=True)
        # if not nltk.data.find("tokenizers/punkt"):  # TODO: check why it is not working
        #     pass

    def _compute(
            self,
            predictions,
            references,
            model_type: Optional[str] = None,
            tokenize_sentences: bool = True,
            multi_references: bool = False,
            gpu: DEVICE_TYPE = False,
            batch_size: int = 32,
            verbose: bool = False,
    ) -> List[Scores]:
        """
            Compute precision, recall, and F1 scores for given predictions and references.

            :param predictions
            :param references
            :param model_type: Type of model to use for encoding.
                Options: [pv1, stsb, use]
                    pv1 - paraphrase-distilroberta-base-v1 (Default)
                    stsb - stsb-roberta-large
                    use - Universal Sentence Encoder
            :param tokenize_sentences: Flag to sentence tokenize the document.
            :param multi_references: Flag to indicate multiple references.
            :param gpu: GPU device to use.
            :param batch_size: Batch size for encoding.
            :param verbose: Flag to indicate verbose output.

            :return: List of Scores dataclass with precision, recall, and F1 scores.
        """

        # Validate inputs corresponding to flags
        _validate_input_format(tokenize_sentences, multi_references, predictions, references)

        # Get GPU
        device = get_gpu(gpu)
        if verbose:
            print(f"Using devices: {device}")

        # Get the encoder model
        model_name = self._get_model_name(model_type)
        encoder = get_encoder(model_name, device=device, batch_size=batch_size, verbose=verbose)

        # We'll handle the single reference and multi-reference case same way. So change the data format accordingly
        if not multi_references:
            references = [[ref] for ref in references]

        # Tokenize sentences if required
        if tokenize_sentences:
            predictions = [nltk.tokenize.sent_tokenize(pred) for pred in predictions]
            references = [[nltk.tokenize.sent_tokenize(ref) for ref in refs] for refs in references]

        # Flatten the data for batch processing
        all_sentences = flatten_list(predictions) + flatten_list(references)

        # Get num of sentences to get the corresponding embeddings
        prediction_sentences_count = [len(pred) for pred in predictions]
        reference_sentences_count = [[len(ref) for ref in refs] for refs in references]

        # Note: This is the most optimal way of doing it
        # Encode all sentences in one go
        embeddings = encoder.encode(all_sentences)

        # Get embeddings corresponding to predictions and references
        pred_embeddings = slice_embeddings(embeddings, prediction_sentences_count)
        ref_embeddings = slice_embeddings(embeddings[sum(prediction_sentences_count):], reference_sentences_count)

        # Init output scores
        results = []

        # Compute scores
        for preds, refs in zip(pred_embeddings, ref_embeddings):
            # Precision: Concatenate all the sentences in all the references
            concat_refs = np.concatenate(refs, axis=0)
            precision, _ = _compute_cosine_similarity(preds, concat_refs)

            # Recall: Compute individually for each reference
            recall_scores = [_compute_cosine_similarity(r_embeds, preds) for r_embeds in refs]
            recall_scores = [r_scores for (r_scores, _) in recall_scores]

            results.append(Scores(precision, recall_scores))

        return results