--- configs: - config_name: fake_news data_files: - split: train path: "fake_news/train.jsonl" - split: test path: "fake_news/test.jsonl" - split: validation path: "fake_news/validation.jsonl" - config_name: job_scams data_files: - split: train path: "job_scams/train.jsonl" - split: test path: "job_scams/test.jsonl" - split: validation path: "job_scams/validation.jsonl" - config_name: phishing data_files: - split: train path: "phishing/train.jsonl" - split: test path: "phishing/test.jsonl" - split: validation path: "phishing/validation.jsonl" - config_name: political_statements data_files: - split: train path: "political_statements/train.jsonl" - split: test path: "political_statements/test.jsonl" - split: validation path: "political_statements/validation.jsonl" - config_name: product_reviews data_files: - split: train path: "product_reviews/train.jsonl" - split: test path: "product_reviews/test.jsonl" - split: validation path: "product_reviews/validation.jsonl" - config_name: sms data_files: - split: train path: "sms/train.jsonl" - split: test path: "sms/test.jsonl" - split: validation path: "sms/validation.jsonl" - config_name: twitter_rumours data_files: - split: train path: "twitter_rumours/train.jsonl" - split: test path: "twitter_rumours/test.jsonl" - split: validation path: "twitter_rumours/validation.jsonl" --- # GDDs-2.0 The Generalized Deception Dataset version 2.0 is a labeled corpus containing over 95000 samples of deceptive and truthful texts from a number of independent domains and tasks. ## Authors ReDAS Lab, University of Houston, 2023. See https://www2.cs.uh.edu/~rmverma/ for contact information. ## Domains/Sub-Tasks There are 7 independent domains in the dataset. Each one was constructed from one or more datasets. Some tasks were not initially binary and had to be relabeled. The inputs vary wildly both stylistically and syntactically, as well as in terms of the goal of deception (or absence of thereof) being performed in the context of each dataset. The two uniting factors are: all seven datasets contain some fraction of texts that are meant to deceive the person reading them one way or another. 1) Phishing (Email) 2) Fake News (News Articles) 3) Political Statements (Claims and statements by politicians and other entities, made from LIAR) 4) Product Reviews (Amazon product reviews) 5) Job Scams (Job postings) 6) SMS (Phishing attacks via sms) (combination of SMS Spam from UCI repository and SMS Phishing datasets) 7) Twitter Rumours (Collection of rumours from PHEME dataset, covers multiple topics) ## Contents Each task is (or has been converted to) a binary classification problem. ### Structure The directory layout of gdds is like so: `` gdds fake_news/ train.jsonl test.jsonl validation.jsonl README.md ... ... ... sms/ train.jsonl test.jsonl validation.jsonl README.md README.md LICENSE.txt `` Each subdirectory/config contains the domain/individual dataset. `train.jsonl`, `test.jsonl`, and `valid.jsonl` contain train, test, and validation sets, respectively. The splits are train=80%, test=10%, valid=10% The sampling process was random with seed=42, and stratified with respect to `y` (label) for each domain. ### Fields Each `jsonl` file has two fields (columns): `text` and `label` `label` answers the question whether text is deceptive: `1` means yes, it is deceptive, `0` means no, the text is not deceptive (it is truthful). `text` is guaranteed to be valid unicode, less than 1 million characters, and contains no empty entries or non-values. ### Documentation Primary documentation is this README file. Each dataset's directory contains a `README.md` file with additional details. The contents of these files are also included at the end of this document in the Appendix. LICENSE.txt contains the MIT license this dataset is distributed under. ## Changes and Additions This dataset is a successor of [the GDD dataset](https://zenodo.org/record/6512468). Noteable changes from GDD are: 1) Addition of SMS and Twitter Rumours datasets, making it 7 deception datasets from different domains in total 2) Re-labeling of Political Statements dataset using a scheme that better fits with prior published work that used it and is stricter in terms of non-deceptive statement criteria of acceptance (see the README file specific to the dataset within its directory) 3) Job Scams datasets' labeles were previously inverted, with ~13500 labeled as deceptive (is_deceptive=True) and ~600 as non-deceptive. This could lead to potential issues with using metrics such as f1-score, which for binary classification is computed for the class considered to be positive. This issue has been addressed and the deceptive texts are labeled as 1 (e.g. positive or True) while non-deceptive as 0 (e.g. negative or False) 4) All datasets have been processed using Cleanlab, with problematic samples maually examined and issues addressed if needed. See the details in each of the individual datasets README files. 5) All datasets now come in 2 formats: the entirety of the data in a single jsonl file located in the `data/` subdirectory of each dataset, and a standard train-test-valid stratified split of 80-10-10, in 3 separate jsonl files. 6) All datasets have two fields: "text" (string) and "label" (integer, 0 or 1 - 0 indicates that the text is non-deceptive, 1 means it is deceptive) 7) '\n' has been normalized to ' ' for all datasets as it causes issues with BERT's tokenizer in some cases (and to be in line with general whitespace normalization). Broken unicode has been fixed. Whitespace, quotations, and bullet points were normalized. Text is limited to 1,000,000 characters in length and guaranteed to be non-empty. Duplicates within the the same dataset (even in text only) were dropped, so were empty and None values. ## Statistics The entire dataset contains 95854 samples, 37282 are deceptive and 58572 non-deceptive. **The split of data within the individual datasets/domains:** fake_news total: 20456 deceptive: 8832 non-deceptive: 11624 job_scams total: 14295 deceptive: 599 non-deceptive: 13696 phishing total: 15272 deceptive: 6074 non-deceptive: 9198 political_statements total: 12497 deceptive: 8042 non-deceptive: 4455 product_reviews total: 20971 deceptive: 10492 non-deceptive: 10479 sms total: 6574 deceptive: 1274 non-deceptive: 5300 twitter_rumours total: 5789 deceptive: 1969 non-deceptive: 3820 ## LICENSE This dataset is published under the MIT license and can be used and modified by anyone free of charge. See LICENSE.txt file for details. ## CITING If you found this dataset useful in your research, please consider citing it as: TODO: ADD our paper reference ## REFERENCES Original GDD paper: @inproceedings{10.1145/3508398.3519358, author = {Zeng, Victor and Liu, Xuting and Verma, Rakesh M.}, title = {Does Deception Leave a Content Independent Stylistic Trace?}, year = {2022}, isbn = {9781450392204}, publisher = {Association for Computing Machinery}, address = {New York, NY, USA}, url = {https://doi.org/10.1145/3508398.3519358}, doi = {10.1145/3508398.3519358}, abstract = {A recent survey claims that there are em no general linguistic cues for deception. Since Internet societies are plagued with deceptive attacks such as phishing and fake news, this claim means that we must build individual datasets and detectors for each kind of attack. It also implies that when a new scam (e.g., Covid) arrives, we must start the whole process of data collection, annotation, and model building from scratch. In this paper, we put this claim to the test by building a quality domain-independent deception dataset and investigating whether a model can perform well on more than one form of deception.}, booktitle = {Proceedings of the Twelfth ACM Conference on Data and Application Security and Privacy}, pages = {349–351}, numpages = {3}, keywords = {domain-independent deception detection, dataset quality/cleaning}, location = {Baltimore, MD, USA}, series = {CODASPY '22} } ## APPENDIX: Dataset and Domain Details This section describes each domain/dataset in greater detail. ### Fake News 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 #### Cleaning 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. #### Preprocessing Whitespace, quotes, bulletpoints, unicode is normalized. #### Data The dataset consists of "text" (string) and "is_deceptive" (1,0). 1 means the text is deceptive, 0 indicates otherwise. 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. ### Job Scams 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 #### Cleaning 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. #### Preprocessing Whitespace, quotes, bulletpoints, unicode is normalized. #### Data The dataset consists of "text" (string) and "is_deceptive" (1,0). 1 means the text is deceptive, 0 indicates otherwise. 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. ### Phishing This dataset consists of various phishing attacks as well as benign emails collected from real users. #### Cleaning 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. #### Preprocessing Whitespace, quotes, bulletpoints, unicode is normalized. #### Data The dataset consists of "text" (string) and "is_deceptive" (1,0). 1 means the text is deceptive, 0 indicates otherwise. 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. ### Political Statements Political Statements dataset was created from the LIAR corpus. #### Labeling The primary difference is the change in the re-labeling scheme when converting the task from multiclass to binary. #### Old scheme We use the claim field as the text and map labels “pants-fire,” “false,” “barely-true,” to deceptive and “half-true,” “mostly-true,” and “true” to non-deceptive, resulting in 5,669 deceptive and 7,167 truthful statements. #### New scheme Following *Upadhayay, B., Behzadan, V.: "Sentimental liar: Extended corpus and deep learning models for fake claim classification" (2020)* and *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.* we map the labels map labels “pants-fire,” “false,” “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. #### Cleaning 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. #### Preprocessing Whitespace, quotes, bulletpoints, unicode is normalized. #### Data The dataset consists of "text" (string) and "is_deceptive" (1,0). 1 means the text is deceptive, 0 indicates otherwise. 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. ### Product Reviews We post-process and split Product Reviews dataset to ensure uniformity with Political Statements 2.0 and Twitter Rumours as they all go into form GDDS-2.0 #### Cleaning 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. #### Preprocessing Whitespace, quotes, bulletpoints, unicode is normalized. #### Data The dataset consists of "text" (string) and "is_deceptive" (1,0). 1 means the text is deceptive, 0 indicates otherwise. 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. ### SMS 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. #### Cleaning 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. #### Preprocessing Whitespace, quotes, bulletpoints, unicode is normalized. #### Data The dataset consists of "text" (string) and "is_deceptive" (1,0). 1 means the text is deceptive, 0 indicates otherwise. 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. ### Rumors dataset This deception dataset was created using PHEME dataset from https://figshare.com/articles/dataset/PHEME_dataset_of_rumours_and_non-rumours/4010619/1 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. #### Cleaning 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. #### Preprocessing Whitespace, quotes, bulletpoints, unicode is normalized. #### Data The dataset consists of "text" (string) and "is_deceptive" (1,0). 1 means the text is deceptive, 0 indicates otherwise. 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.