dainis-boumber commited on
Commit
a7a8c3b
1 Parent(s): 64b9954

updated dataset descriptions

Browse files
Files changed (1) hide show
  1. README.md +65 -21
README.md CHANGED
@@ -219,19 +219,23 @@ series = {CODASPY '22}
219
  }
220
 
221
 
222
-
223
-
224
  ## APPENDIX: Dataset and Domain Details
225
 
226
  This section describes each domain/dataset in greater detail.
227
 
228
  ### Fake News
229
 
230
- We post-process and split Fake News dataset to ensure uniformity with Political Statements 2.0 and Twitter Rumours as they all go into form GDDS-2.0
 
 
 
 
 
231
 
232
  #### Cleaning
233
 
234
- Each dataset has been cleaned using Cleanlab. Non-english entries, erroneous (parser error) entries, empty entries, duplicate entries, entries of length less than 2 characters or exceeding 1000000 characters were all removed.
 
235
 
236
  #### Preprocessing
237
 
@@ -241,15 +245,25 @@ Whitespace, quotes, bulletpoints, unicode is normalized.
241
 
242
  The dataset consists of "text" (string) and "is_deceptive" (1,0). 1 means the text is deceptive, 0 indicates otherwise.
243
 
244
- There are 20456 samples in the dataset, contained in `phishing.jsonl`. For reproduceability, the data is also split into training, test, and validation sets in 80/10/10 ratio. They are named `train.jsonl`, `test.jsonl`, `valid.jsonl`. The sampling process was stratified. The training set contains 16364 samples, the validation and the test sets have 2064 and 2064 samles, respectively.
 
 
245
 
246
  ### Job Scams
247
 
248
- We post-process and split Job Scams dataset to ensure uniformity with Political Statements 2.0 and Twitter Rumours as they all go into form GDDS-2.0
 
 
 
 
 
249
 
250
  #### Cleaning
251
 
252
- Each dataset has been cleaned using Cleanlab. Non-english entries, erroneous (parser error) entries, empty entries, duplicate entries, entries of length less than 2 characters or exceeding 1000000 characters were all removed.
 
 
 
253
 
254
  #### Preprocessing
255
 
@@ -259,7 +273,10 @@ Whitespace, quotes, bulletpoints, unicode is normalized.
259
 
260
  The dataset consists of "text" (string) and "is_deceptive" (1,0). 1 means the text is deceptive, 0 indicates otherwise.
261
 
262
- There are 14295 samples in the dataset, contained in `job_scams.jsonl`. For reproduceability, the data is also split into training, test, and validation sets in 80/10/10 ratio. They are named `train.jsonl`, `test.jsonl`, `valid.jsonl`. The sampling process was stratified. The training set contains 11436 samples, the validation and the test sets have 1429 and 1430 samles, respectively.
 
 
 
263
 
264
  ### Phishing
265
 
@@ -267,7 +284,8 @@ This dataset consists of various phishing attacks as well as benign emails colle
267
 
268
  #### Cleaning
269
 
270
- Each dataset has been cleaned using Cleanlab. Non-english entries, erroneous (parser error) entries, empty entries, duplicate entries, entries of length less than 2 characters or exceeding 1000000 characters were all removed.
 
271
 
272
  #### Preprocessing
273
 
@@ -277,11 +295,15 @@ Whitespace, quotes, bulletpoints, unicode is normalized.
277
 
278
  The dataset consists of "text" (string) and "is_deceptive" (1,0). 1 means the text is deceptive, 0 indicates otherwise.
279
 
280
- There are 15272 samples in the dataset, contained in `phishing.jsonl`. For reproduceability, the data is also split into training, test, and validation sets in 80/10/10 ratio. They are named `train.jsonl`, `test.jsonl`, `valid.jsonl`. The sampling process was stratified. The training set contains 12217 samples, the validation and the test sets have 1527 and 1528 samles, respectively.
 
 
 
281
 
282
  ### Political Statements
283
 
284
- Political Statements dataset was created from the LIAR corpus.
 
285
 
286
  #### Labeling
287
 
@@ -305,11 +327,16 @@ and
305
  *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.*
306
 
307
  we map the labels map labels “pants-fire,” “false,”
308
- “barely-true,” **and “half-true,”** to deceptive; the labels "mostly-true" and "true" are mapped to non-deceptive. The statements that are only half-true are now considered to be deceptive, making the criterion for statement being non-deceptive stricter -- now 2 out of 6 labels map to non-deceptive and 4 map to deceptive.
 
 
309
 
310
  #### Cleaning
311
 
312
- The dataset has been cleaned using cleanlab with visual inspection of problems found. Partial sentences, such as "On Iran nuclear deal", "On inflation", were removed. Text with large number of errors induced by a parser were also removed. Statements in language other than English (namely, Spanish) were also removed. Sequences with unicode errors, containing less than one characters or over 1 million characters were removed.
 
 
 
313
 
314
  #### Preprocessing
315
 
@@ -319,7 +346,10 @@ Whitespace, quotes, bulletpoints, unicode is normalized.
319
 
320
  The dataset consists of "text" (string) and "is_deceptive" (1,0). 1 means the text is deceptive, 0 indicates otherwise.
321
 
322
- There are 12497 samples in the dataset, contained in `political_statements.jsonl`. For reproduceability, the data is also split into training, test, and validation sets in 80/10/10 ratio. They are named `train.jsonl`, `test.jsonl`, `valid.jsonl`. The sampling process was stratified. The training set contains 9997 samples, the validation and the test sets have 1250 samles each in them.
 
 
 
323
 
324
  ### Product Reviews
325
 
@@ -337,15 +367,22 @@ Whitespace, quotes, bulletpoints, unicode is normalized.
337
 
338
  The dataset consists of "text" (string) and "is_deceptive" (1,0). 1 means the text is deceptive, 0 indicates otherwise.
339
 
340
- There are 20971 samples in the dataset, contained in `product_reviews.jsonl`. For reproduceability, the data is also split into training, test, and validation sets in 80/10/10 ratio. They are named `train.jsonl`, `test.jsonl`, `valid.jsonl`. The sampling process was stratified. The training set contains 16776 samples, the validation and the test sets have 2097 and 2098 samles, respectively.
 
 
 
341
 
342
  ### SMS
343
 
344
- This dataset was created from the SMS Spam Collection and SMS Phishing Dataset for Machine Learning and Pattern Recognition, which contained 5,574 and 5,971 real English SMS messages, respectively. As these two datasets overlap, after de-duplication, the final dataset is made up of 6574 texts released by a private UK-based wireless operator; 1274 of them are deceptive, and the remaining 5300 are not.
 
 
 
345
 
346
  #### Cleaning
347
 
348
- Each dataset has been cleaned using Cleanlab. Non-english entries, erroneous (parser error) entries, empty entries, duplicate entries, entries of length less than 2 characters or exceeding 1000000 characters were all removed.
 
349
 
350
  #### Preprocessing
351
 
@@ -355,7 +392,10 @@ Whitespace, quotes, bulletpoints, unicode is normalized.
355
 
356
  The dataset consists of "text" (string) and "is_deceptive" (1,0). 1 means the text is deceptive, 0 indicates otherwise.
357
 
358
- There are 6574 samples in the dataset, contained in `sms.jsonl`. For reproduceability, the data is also split into training, test, and validation sets in 80/10/10 ratio. They are named `train.jsonl`, `test.jsonl`, `valid.jsonl`. The sampling process was stratified. The training set contains 5259 samples, the validation and the test sets have 657 and 658 samles, respectively.
 
 
 
359
 
360
  ### Rumors dataset
361
 
@@ -363,11 +403,13 @@ This deception dataset was created using PHEME dataset from
363
 
364
  https://figshare.com/articles/dataset/PHEME_dataset_of_rumours_and_non-rumours/4010619/1
365
 
366
- was used in creation of this dataset. We took source tweets only, and ignored replies to them. We used source tweet's label as being a rumour or non-rumour to label it as deceptive or non-deceptive.
 
367
 
368
  #### Cleaning
369
 
370
- The dataset has been cleaned using cleanlab with visual inspection of problems found. No issues were identified. Duplicate entries, entries of length less than 2 characters or exceeding 1000000 characters were removed.
 
371
 
372
  #### Preprocessing
373
 
@@ -377,7 +419,9 @@ Whitespace, quotes, bulletpoints, unicode is normalized.
377
 
378
  The dataset consists of "text" (string) and "is_deceptive" (1,0). 1 means the text is deceptive, 0 indicates otherwise.
379
 
380
- There are 5789 samples in the dataset, contained in `tweeter_rumours.jsonl`. For reproduceability, the data is also split into training, test, and validation sets in 80/10/10 ratio. They are named `train.jsonl`, `test.jsonl`, `valid.jsonl`. The sampling process was stratified. The training set contains 4631 samples, the validation and the test sets have 579 samles each.
 
 
381
 
382
 
383
 
 
219
  }
220
 
221
 
 
 
222
  ## APPENDIX: Dataset and Domain Details
223
 
224
  This section describes each domain/dataset in greater detail.
225
 
226
  ### Fake News
227
 
228
+ Fake News used WELFake as a basis. The WELFake dataset combines 72,134 news articles from four pre-existing datasets
229
+ (Kaggle, McIntire, Reuters, and BuzzFeed Political). The dataset was cleaned of data leaks in the form of citations of
230
+ often reputable sources, such as "[claim] (Reuters)". It contains 35,028 real news articles and 37,106 fake news articles.
231
+ We found a number of out-of-domain statements that are clearly not relevant to news, such as "Cool", which is a potential
232
+ problem for transfer learning as well as classification. After cleaning and processing, the Fake News dataset consists of
233
+ 20456 articles; 8832 are deceptive, and 11624 are not.
234
 
235
  #### Cleaning
236
 
237
+ Each dataset has been cleaned using Cleanlab. Non-english entries, erroneous (parser error) entries, empty entries, duplicate entries,
238
+ entries of length less than 2 characters or exceeding 1000000 characters were all removed.
239
 
240
  #### Preprocessing
241
 
 
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
253
 
254
+ The Employment Scam Aegean Dataset, henceforth referred to as the Job Scams dataset, consisted of 17,880 human-annotated job listings of
255
+ job descriptions labeled as fraudulent or not.
256
+
257
+ #### Relabeling
258
+
259
+ The original Job Labels dataset had the labels inverted when released. The problem is now fixed, the labels are correct.
260
 
261
  #### Cleaning
262
 
263
+ It was cleaned by removing all HTML tags, empty descriptions, and duplicates. The dataset has been cleaned using Cleanlab.
264
+ Non-english entries, erroneous (parser error) entries, empty entries, duplicate entries, entries of length less
265
+ than 2 characters or exceeding 1000000 characters were all removed.
266
+ The final dataset is heavily imbalanced, with 599 deceptive and 13696 non-deceptive samples out of the 14295 total.
267
 
268
  #### Preprocessing
269
 
 
273
 
274
  The dataset consists of "text" (string) and "is_deceptive" (1,0). 1 means the text is deceptive, 0 indicates otherwise.
275
 
276
+ There are 14295 samples in the dataset, contained in `job_scams.jsonl`.
277
+ For reproduceability, the data is also split into training, test, and validation sets in 80/10/10 ratio.
278
+ They are named `train.jsonl`, `test.jsonl`, `valid.jsonl`. The sampling process was stratified.
279
+ The training set contains 11436 samples, the validation and the test sets have 1429 and 1430 samles, respectively.
280
 
281
  ### Phishing
282
 
 
284
 
285
  #### Cleaning
286
 
287
+ Each dataset has been cleaned using Cleanlab. Non-english entries, erroneous (parser error) entries, empty entries,
288
+ duplicate entries, entries of length less than 2 characters or exceeding 1000000 characters were all removed.
289
 
290
  #### Preprocessing
291
 
 
295
 
296
  The dataset consists of "text" (string) and "is_deceptive" (1,0). 1 means the text is deceptive, 0 indicates otherwise.
297
 
298
+ There are 15272 samples in the dataset, contained in `phishing.jsonl`.
299
+ For reproduceability, the data is also split into training, test, and validation sets in 80/10/10 ratio.
300
+ They are named `train.jsonl`, `test.jsonl`, `valid.jsonl`. The sampling process was stratified.
301
+ The training set contains 12217 samples, the validation and the test sets have 1527 and 1528 samles, respectively.
302
 
303
  ### Political Statements
304
 
305
+ This corpus was created from the Liar dataset which consists of political statements made by US speakers assigned
306
+ a fine-grain truthfulness label by PolitiFact.
307
 
308
  #### Labeling
309
 
 
327
  *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.*
328
 
329
  we map the labels map labels “pants-fire,” “false,”
330
+ “barely-true,” **and “half-true,”** to deceptive; the labels "mostly-true" and "true" are mapped to non-deceptive.
331
+ The statements that are only half-true are now considered to be deceptive, making the criterion for statement being non-deceptive stricter:
332
+ now 2 out of 6 labels map to non-deceptive and 4 map to deceptive.
333
 
334
  #### Cleaning
335
 
336
+ The dataset has been cleaned using cleanlab with visual inspection of problems found. Partial sentences, such as "On Iran nuclear deal",
337
+ "On inflation", were removed. Text with large number of errors induced by a parser were also removed.
338
+ Statements in language other than English (namely, Spanish) were also removed.
339
+ Sequences with unicode errors, containing less than one characters or over 1 million characters were removed.
340
 
341
  #### Preprocessing
342
 
 
346
 
347
  The dataset consists of "text" (string) and "is_deceptive" (1,0). 1 means the text is deceptive, 0 indicates otherwise.
348
 
349
+ There are 12497 samples in the dataset, contained in `political_statements.jsonl`.
350
+ For reproduceability, the data is also split into training, test, and validation sets in 80/10/10 ratio.
351
+ They are named `train.jsonl`, `test.jsonl`, `valid.jsonl`. The sampling process was stratified.
352
+ The training set contains 9997 samples, the validation and the test sets have 1250 samles each in them.
353
 
354
  ### Product Reviews
355
 
 
367
 
368
  The dataset consists of "text" (string) and "is_deceptive" (1,0). 1 means the text is deceptive, 0 indicates otherwise.
369
 
370
+ There are 20971 samples in the dataset, contained in `product_reviews.jsonl`.
371
+ For reproduceability, the data is also split into training, test, and validation sets in 80/10/10 ratio.
372
+ They are named `train.jsonl`, `test.jsonl`, `valid.jsonl`. The sampling process was stratified.
373
+ The training set contains 16776 samples, the validation and the test sets have 2097 and 2098 samles, respectively.
374
 
375
  ### SMS
376
 
377
+ This dataset was created from the SMS Spam Collection and SMS Phishing Dataset for Machine Learning and Pattern Recognition,
378
+ which contained 5,574 and 5,971 real English SMS messages, respectively. As these two datasets overlap, after de-duplication,
379
+ the final dataset is made up of 6574 texts released by a private UK-based wireless operator; 1274 of them are deceptive,
380
+ and the remaining 5300 are not.
381
 
382
  #### Cleaning
383
 
384
+ Each dataset has been cleaned using Cleanlab. Non-english entries, erroneous (parser error) entries, empty entries,
385
+ duplicate entries, entries of length less than 2 characters or exceeding 1000000 characters were all removed.
386
 
387
  #### Preprocessing
388
 
 
392
 
393
  The dataset consists of "text" (string) and "is_deceptive" (1,0). 1 means the text is deceptive, 0 indicates otherwise.
394
 
395
+ There are 6574 samples in the dataset, contained in `sms.jsonl`. For reproduceability, the data is also split into training,
396
+ test, and validation sets in 80/10/10 ratio. They are named `train.jsonl`, `test.jsonl`, `valid.jsonl`.
397
+ The sampling process was stratified. The training set contains 5259 samples, the validation and the test sets have 657 and 658 samles,
398
+ respectively.
399
 
400
  ### Rumors dataset
401
 
 
403
 
404
  https://figshare.com/articles/dataset/PHEME_dataset_of_rumours_and_non-rumours/4010619/1
405
 
406
+ was used in creation of this dataset. We took source tweets only, and ignored replies to them.
407
+ We used source tweet's label as being a rumour or non-rumour to label it as deceptive or non-deceptive.
408
 
409
  #### Cleaning
410
 
411
+ The dataset has been cleaned using cleanlab with visual inspection of problems found. No issues were identified.
412
+ Duplicate entries, entries of length less than 2 characters or exceeding 1000000 characters were removed.
413
 
414
  #### Preprocessing
415
 
 
419
 
420
  The dataset consists of "text" (string) and "is_deceptive" (1,0). 1 means the text is deceptive, 0 indicates otherwise.
421
 
422
+ There are 5789 samples in the dataset, contained in `tweeter_rumours.jsonl`. For reproduceability, the data is also split into training,
423
+ test, and validation sets in 80/10/10 ratio. They are named `train.jsonl`, `test.jsonl`, `valid.jsonl`.
424
+ The sampling process was stratified. The training set contains 4631 samples, the validation and the test sets have 579 samles each.
425
 
426
 
427