dainis-boumber commited on
Commit
3b77f86
1 Parent(s): 45bc599
Files changed (1) hide show
  1. README.md +40 -60
README.md CHANGED
@@ -3,59 +3,66 @@ configs:
3
  - config_name: fake_news
4
  data_files:
5
  - split: train
6
- path: "fake_news/train.jsonl"
7
  - split: test
8
- path: "fake_news/test.jsonl"
9
  - split: validation
10
- path: "fake_news/validation.jsonl"
11
  - config_name: job_scams
12
  data_files:
13
  - split: train
14
- path: "job_scams/train.jsonl"
15
  - split: test
16
- path: "job_scams/test.jsonl"
17
  - split: validation
18
- path: "job_scams/validation.jsonl"
19
  - config_name: phishing
20
  data_files:
21
  - split: train
22
- path: "phishing/train.jsonl"
23
  - split: test
24
- path: "phishing/test.jsonl"
25
  - split: validation
26
- path: "phishing/validation.jsonl"
27
  - config_name: political_statements
28
  data_files:
29
  - split: train
30
- path: "political_statements/train.jsonl"
31
  - split: test
32
- path: "political_statements/test.jsonl"
33
  - split: validation
34
- path: "political_statements/validation.jsonl"
35
  - config_name: product_reviews
36
  data_files:
37
  - split: train
38
- path: "product_reviews/train.jsonl"
39
  - split: test
40
- path: "product_reviews/test.jsonl"
41
  - split: validation
42
- path: "product_reviews/validation.jsonl"
43
  - config_name: sms
44
  data_files:
45
  - split: train
46
- path: "sms/train.jsonl"
47
  - split: test
48
- path: "sms/test.jsonl"
49
  - split: validation
50
- path: "sms/validation.jsonl"
51
  - config_name: twitter_rumours
52
  data_files:
53
  - split: train
54
- path: "twitter_rumours/train.jsonl"
55
  - split: test
56
- path: "twitter_rumours/test.jsonl"
57
  - split: validation
58
- path: "twitter_rumours/validation.jsonl"
 
 
 
 
 
 
 
59
  ---
60
 
61
  # GDDs-2.0
@@ -227,7 +234,6 @@ location = {Baltimore, MD, USA},
227
  series = {CODASPY '22}
228
  }
229
 
230
-
231
  ## APPENDIX: Dataset and Domain Details
232
 
233
  This section describes each domain/dataset in greater detail.
@@ -240,13 +246,6 @@ often reputable sources, such as "[claim] (Reuters)". It contains 35,028 real ne
240
  We found a number of out-of-domain statements that are clearly not relevant to news, such as "Cool", which is a potential
241
  problem for transfer learning as well as classification.
242
 
243
-
244
- #### Data
245
-
246
- The dataset consists of "text" (string) and "is_deceptive" (1,0). 1 means the text is deceptive, 0 indicates otherwise.
247
-
248
- There are 20456 samples in the dataset, contained in `phishing.jsonl`. For reproduceability, the data is also split into training, test,
249
- and validation sets in 80/10/10 ratio. They are named `train.jsonl`, `test.jsonl`, `valid.jsonl`. The sampling process was stratified.
250
  The training set contains 16364 samples, the validation and the test sets have 2064 and 2064 samles, respectively.
251
 
252
  ### JOB SCAMS
@@ -260,18 +259,13 @@ The original Job Labels dataset had the labels inverted when released. The probl
260
 
261
  #### Cleaning
262
 
263
- HTML tags were removed.
264
-
265
- #### Data
266
-
267
- T**With just under 600 deceptive texts, this dataset is heavily imbalanced.**
268
 
269
  ### PHISHING
270
 
271
  This dataset consists of various phishing attacks as well as benign emails collected from real users.
272
 
273
- #### Data
274
-
275
  The training set contains 12217 samples, the validation and the test sets have 1527 and 1528 samples, respectively.
276
 
277
  ### POLITICAL STATEMENTS
@@ -298,7 +292,8 @@ Following
298
 
299
  and
300
 
301
- *Shahriar, Sadat, Arjun Mukherjee, and Omprakash Gnawali. "Deception Detection with Feature-Augmentation by Soft Domain Transfer." International Conference on Social Informatics. Cham: Springer International Publishing, 2022.*
 
302
 
303
  we map the labels map labels “pants-fire,” “false,”
304
  “barely-true,” **and “half-true,”** to deceptive; the labels "mostly-true" and "true" are mapped to non-deceptive.
@@ -311,8 +306,6 @@ The dataset has been cleaned using cleanlab with visual inspection of problems f
311
  "On inflation", were removed. Text with large number of errors induced by a parser were also removed.
312
  Statements in language other than English (namely, Spanish) were also removed.
313
 
314
- #### Data
315
-
316
  The training set contains 9997 samples, the validation and the test sets have 1250 samples each in them.
317
 
318
  ### PRODUCT REVIEWS
@@ -320,7 +313,13 @@ The training set contains 9997 samples, the validation and the test sets have 12
320
  We post-process and split Product Reviews dataset to ensure uniformity with Political Statements 2.0 and Twitter Rumours
321
  as they all go into form GDDS-2.0
322
 
323
- #### Data
 
 
 
 
 
 
324
 
325
  The training set contains 16776 samples, the validation and the test sets have 2097 and 2098 samples, respectively.
326
 
@@ -331,8 +330,6 @@ which contained 5,574 and 5,971 real English SMS messages, respectively. As thes
331
  the final dataset is made up of 6574 texts released by a private UK-based wireless operator; 1274 of them are deceptive,
332
  and the remaining 5300 are not.
333
 
334
- #### Data
335
-
336
  The training set contains 5259 samples, the validation and the test sets have 657 and 658 samples,
337
  respectively.
338
 
@@ -345,21 +342,4 @@ https://figshare.com/articles/dataset/PHEME_dataset_of_rumours_and_non-rumours/4
345
  was used in creation of this dataset. We took source tweets only, and ignored replies to them.
346
  We used source tweet's label as being a rumour or non-rumour to label it as deceptive or non-deceptive.
347
 
348
- #### Data
349
-
350
- The training set contains 4631 samples, the validation and the test sets have 579 samples each.
351
-
352
-
353
-
354
-
355
-
356
-
357
-
358
-
359
-
360
-
361
-
362
-
363
-
364
-
365
-
 
3
  - config_name: fake_news
4
  data_files:
5
  - split: train
6
+ path: fake_news/train.jsonl
7
  - split: test
8
+ path: fake_news/test.jsonl
9
  - split: validation
10
+ path: fake_news/validation.jsonl
11
  - config_name: job_scams
12
  data_files:
13
  - split: train
14
+ path: job_scams/train.jsonl
15
  - split: test
16
+ path: job_scams/test.jsonl
17
  - split: validation
18
+ path: job_scams/validation.jsonl
19
  - config_name: phishing
20
  data_files:
21
  - split: train
22
+ path: phishing/train.jsonl
23
  - split: test
24
+ path: phishing/test.jsonl
25
  - split: validation
26
+ path: phishing/validation.jsonl
27
  - config_name: political_statements
28
  data_files:
29
  - split: train
30
+ path: political_statements/train.jsonl
31
  - split: test
32
+ path: political_statements/test.jsonl
33
  - split: validation
34
+ path: political_statements/validation.jsonl
35
  - config_name: product_reviews
36
  data_files:
37
  - split: train
38
+ path: product_reviews/train.jsonl
39
  - split: test
40
+ path: product_reviews/test.jsonl
41
  - split: validation
42
+ path: product_reviews/validation.jsonl
43
  - config_name: sms
44
  data_files:
45
  - split: train
46
+ path: sms/train.jsonl
47
  - split: test
48
+ path: sms/test.jsonl
49
  - split: validation
50
+ path: sms/validation.jsonl
51
  - config_name: twitter_rumours
52
  data_files:
53
  - split: train
54
+ path: twitter_rumours/train.jsonl
55
  - split: test
56
+ path: twitter_rumours/test.jsonl
57
  - split: validation
58
+ path: twitter_rumours/validation.jsonl
59
+ license: mit
60
+ task_categories:
61
+ - text-classification
62
+ language:
63
+ - en
64
+ size_categories:
65
+ - 10K<n<100K
66
  ---
67
 
68
  # GDDs-2.0
 
234
  series = {CODASPY '22}
235
  }
236
 
 
237
  ## APPENDIX: Dataset and Domain Details
238
 
239
  This section describes each domain/dataset in greater detail.
 
246
  We found a number of out-of-domain statements that are clearly not relevant to news, such as "Cool", which is a potential
247
  problem for transfer learning as well as classification.
248
 
 
 
 
 
 
 
 
249
  The training set contains 16364 samples, the validation and the test sets have 2064 and 2064 samles, respectively.
250
 
251
  ### JOB SCAMS
 
259
 
260
  #### Cleaning
261
 
262
+ It was cleaned by removing all HTML tags, empty descriptions, and duplicates.
263
+ The final dataset is heavily imbalanced, with 599 deceptive and 13696 non-deceptive samples out of the 14295 total.
 
 
 
264
 
265
  ### PHISHING
266
 
267
  This dataset consists of various phishing attacks as well as benign emails collected from real users.
268
 
 
 
269
  The training set contains 12217 samples, the validation and the test sets have 1527 and 1528 samples, respectively.
270
 
271
  ### POLITICAL STATEMENTS
 
292
 
293
  and
294
 
295
+ *Shahriar, Sadat, Arjun Mukherjee, and Omprakash Gnawali. "Deception Detection with Feature-Augmentation by Soft Domain Transfer."
296
+ International Conference on Social Informatics. Cham: Springer International Publishing, 2022.*
297
 
298
  we map the labels map labels “pants-fire,” “false,”
299
  “barely-true,” **and “half-true,”** to deceptive; the labels "mostly-true" and "true" are mapped to non-deceptive.
 
306
  "On inflation", were removed. Text with large number of errors induced by a parser were also removed.
307
  Statements in language other than English (namely, Spanish) were also removed.
308
 
 
 
309
  The training set contains 9997 samples, the validation and the test sets have 1250 samples each in them.
310
 
311
  ### PRODUCT REVIEWS
 
313
  We post-process and split Product Reviews dataset to ensure uniformity with Political Statements 2.0 and Twitter Rumours
314
  as they all go into form GDDS-2.0
315
 
316
+ The dataset is produced from English Amazon Reviews labeled as either real or fake, relabeled as deceptive and non-deceptive respectively.
317
+ The reviews cover a variety of products with no particular product dominating the dataset. Although the dataset authors filtered out
318
+ non-English reviews, through outlier detection we found that the dataset still contains reviews in Spanish and other languages.
319
+ Problematic label detection shows that over 6713 samples are potentially mislabeled; since this technique is error-prone,
320
+ we visually examine 67 reviews that are found to be the largest potential sources of error (the top percentile) and confirm that
321
+ most of them appear to be mislabeled. The final dataset of 20,971 reviews is evenly balanced with 10,492 deceptive and 10,479
322
+ non-deceptive samples.
323
 
324
  The training set contains 16776 samples, the validation and the test sets have 2097 and 2098 samples, respectively.
325
 
 
330
  the final dataset is made up of 6574 texts released by a private UK-based wireless operator; 1274 of them are deceptive,
331
  and the remaining 5300 are not.
332
 
 
 
333
  The training set contains 5259 samples, the validation and the test sets have 657 and 658 samples,
334
  respectively.
335
 
 
342
  was used in creation of this dataset. We took source tweets only, and ignored replies to them.
343
  We used source tweet's label as being a rumour or non-rumour to label it as deceptive or non-deceptive.
344
 
345
+ The training set contains 4631 samples, the validation and the test sets have 579 samples each.