Identification of offensive language in Urdu using semantic and embedding models

Framework methodology

Social media text is usually in unstructured form and has a wide variety of visualization formats depending upon the specific platform. It is very hard to observe the offensive language in Urdu using only char and word n-gram feature models. Therefore, we use TF-IDF, bag of words, and word2vec feature extraction models in comparison with word n-gram and char n-gram feature models. The pipeline of the proposed offensive language detection model in the Urdu language is presented in Fig. 1. It takes annotated dataset (posts/comments from Pakistani public pages of Facebook) as input, and preprocesses it by removing punctuation marks, white spaces, accents, and inconsistencies from Urdu text. It then tokenizes Urdu text to further prepare it for feature extraction. After tokenization and stop words removal, features are extracted using word unigram, bag of words, TF-IDF, and word2vec extraction methods. Then state-of-the-art classifiers, evaluation metrics, and a 10-fold cross-validation method are used for experimentation. The outcome of the framework is the binary label (offensive or not offensive) of the post.

Problem formulation

Offensive language detection is the binary classification problem, formally described as follows.

Suppose we have a collection of Facebook posts (P₁, P₂, …P_n) and their corresponding vectors $\left({\mathrm{X}}_1,{\mathrm{Z}}_1\right),\left({\mathrm{X}}_2,{\mathrm{Z}}_2\right),\dots \left({\mathrm{X}}_{\mathrm{n}},{\mathrm{Z}}_{\mathrm{n}}\right)$ The variable n represents the total number of posts; X_i is the feature vector related to the post P_i(X_i ∈ R^T, R^Trefers to the total number of features) and Z_i ∈ offensive, not offensive. To classify whether a Facebook post P_i is offensive or not, a predictive function is defined as follows. (1)${\displaystyle {\mathrm{Z}}_{\mathrm{i}}={\mathrm{F}}_{\mathrm{OL}}\left(\frac{{\mathrm{P}}_{\mathrm{i}}}{{\mathrm{X}}_{\mathrm{i}}}\right).}$

Where (2)${\displaystyle {\mathrm{F}}_{\mathrm{OL}}\left(\frac{{\mathrm{P}}_{\mathrm{i}}}{{\mathrm{X}}_{\mathrm{i}}}\right)=\left[\begin{array}{c}\hfill >\mathrm{0},\mathrm{if}{\mathrm{Z}}_{\mathrm{i}}=\mathrm{1},\text{Offensive}\hfill \\ \hfill =\mathrm{0},\mathrm{if}{\mathrm{Z}}_{\mathrm{i}}=\mathrm{0},\mathrm{Not}-\text{Offensive}\hfill \end{array}\right].}$

Objective function: Our goal is to learn a predictive function that helps to predict whether a post is offensive or not so that future instances can be classified correctly.

Dataset preparation

Here, we present the process of data collection and annotation to create the offensive language dataset in Urdu, and describe the statistics of the resulting dataset.

Domain selection & data collection

We use Facebook’s graph application programming interface (API) to collect posts/comments from Facebook pages. To build a dictionary of seed words, initially, a manual list of offensive words in Urdu is designed. This list is then used to search for other words and keywords used in Facebook posts as offensive. After searching for posts containing these words, the Facebook posts were manually inspected, and more phrases and words are identified. This ended up with enough keywords/words being used as offensive. Some keywords contain more than one word, and some contain only one word. It is necessary to disclose that the selection of these words does not relate to their frequency of occurrence in a post. If a word appears once in a Facebook post to offend someone, it is included in the dictionary such as:

• (Rhinoceros mouth)

• (Shameless)

• (Rubbish)

These are examples of offensive keywords. Popular and diverse Facebook pages from different Pakistani newsgroups (religious groups, political parties groups, and popular bloggers) are selected to build the data corpus. We targeted the most popular Facebook pages because these pages disseminate public opinions rapidly. Likewise, the diversity of source pages makes our dataset a good representative of different categories, e.g., religion, politics, etc. The sampling criteria and metrics used to select a public page from chosen categories are given below:

1. The number of followers and likes should be greater than 30,000, allowing more active public pages to be selected from categories.

2. Only those pages are selected that employ the Urdu language most frequently for posts and comments.

By employing the above criteria, we selected 19 Facebook public pages as described in Table 2. Using the seed word dictionary, we collected Facebook posts/comments containing any of these keywords for 36 months ranging from June 01, 2017, to May 30, 2020. The reason why we choose this period was the general elections held in 2018 in Pakistan and other religious activities. The process of preparation of the dataset with annotation is presented in Fig. 2. Initially, the collection of Facebook posts led to 32,480 instances containing at least one dictionary word. After that, the data was considered for cleaning using the following four steps: (1) since our focus is on the Urdu language, therefore we removed all the posts and comments which are in English, and Roman Urdu, (2) after that, we removed all the punctuation marks, null values, URLs, Emojis, special symbols and numbers from Urdu post/comments because they have no contribution in offensive language detection, (3) we performed normalization on Urdu text to convert homophone variations of Urdu writings to a common symbol e.g., characters ‘ ’ and ‘ ’ are to be replaced by ‘ ’ and also removes spaces and duplication in the text, and (4) using the Urdu stop word list, we removed stop words from our corpus. After the cleaning steps, the dataset finally led to 12,416 posts. This dataset is further considered for annotation.

Table 2:

Selected Facebook public pages.

S #	Facebook pages	Before preprocessing		After preprocessing
		# Of Posts	# Of Comments	# Of Posts	# Of Comments
1	Dunya News	1,000	150,000	460	12,000
2	Samma TV	300	30,000	167	6,600
3	Express News	1,200	100,000	780	22,000
4	Bol News	450	20,000	120	2,500
5	ARY News	800	60,000	333	12,000
6	GeoUrduNews	8,000	105,000	3,000	28,000
7	Urdu Point	2,300	40,000	550	7,500
8	Daily Pakistan	2,800	80,000	880	18,000
9	bbcurdu	5,000	100,000	900	19,000
10	ZemTv	300	18,000	80	1,200
11	PTIOfficial	7,000	110,000	4,400	35,000
12	PPPOfficial	800	45,000	130	8,000
13	PML N Official	1,600	70,000	400	24,500
14	Pervaiz Hood	50	550	10	68
15	JIPOfficial	80	1,000	23	133
16	JUIPK	300	30,000	78	2,500
17	Chingari	400	8,000	80	550
18	marvisirmed	20	1,000	5	125
19	Liberal2016	80	3,500	20	250
	Total	32,480	858,550	12,416	196,226

DOI: 10.7717/peerjcs.1169/table-2

Features extraction

Feature extraction is the most important step in any natural language processing task (NLP). It has been observed that whenever irrelevant features are used, then it may lead to misclassification. Therefore, to design an effective offensive language detection model for the Urdu language, we consider the following state-of-the-art features:

Classifiers and evaluation measures

In this study, five ML algorithms have been selected for experimental setup, to develop a robust model for offensive language detection in Urdu. The models are logistic regression (Logistic-Reg), random forest (RF), stochastic gradient descent (SGD), support vector machine (SVM), and an ensemble model. The pipeline of the ensemble model is presented in Fig. 3. The majority voting methodology is adopted in the ensemble model and Logistic-Reg, SVM, and SGD are the three ML models being used. We compare the performance of five ML models, and the best model is defined.

In addition, the 10-fold cross-validation method is used for model training and testing. The results have been reported using standard accuracy, precision, recall, F1-score, area under the curve (AUC), and Matthews correlation coefficient (MCC) measures.

The mathematical definitions of these metrics are described as follows:

Accuracy

It is measured as the ratio of the number of correctly predicted instances (positive and negative) to all predictions. (3)${\displaystyle \text{Accuracy}=\frac{\text{TruePositive}+\text{TrueNegative}}{\text{TruePositive}+\text{TrueNegative}+\text{FalsePositve}+\text{FalseNegative}}.}$

Precision

It is the measure that summarizes the fraction of actual instances of an offensive class to the total number of instances assigned an offensive class label. (4)${\displaystyle \text{Precision}=\frac{\text{TruePositive}}{\text{TruePositive}+\text{FalsePositive}}.}$

Recall

It summarizes how well the offensive class is predicted and it is calculated as (5)${\displaystyle \text{Recall}=\frac{\text{TruePositive}}{\text{TruePositive}+\text{FalseNegative}}.}$

F1-Score

F1-score is the combination of precision and recall metrics that balances both measures. It is calculated as (6)${\displaystyle \mathrm{F}1-\text{measure}=\frac{2\text{*Precision*Recall}}{\text{Preceision}+\text{Recall}}.}$

The AUC

It relates the true positive rate to the false positive rate and provides an aggregate measure of performance across all possible classification thresholds.

The MCC

It is a reliable statistical rate that produces a high score only if the prediction obtained good results in all of the four confusion matrix categories, i.e., true positives, false negatives, true negatives, and false positives.

Experiment 1: performance comparison of Ml methods and feature models

Here experiments are conducted to meet two objectives: (1) performance comparison of five ML models using proposed features, (2) investigation of the impact of four types of features for offensive language detection in Urdu. The results are evaluated using accuracy, precision, recall, F1-measure, AUC, ROC curves, and MCC. The ML methods are trained and tested using the standard 10-fold cross-validation technique. The performance of each classifier with word unigram, TF-IDF, bag-of-word, and word2vec methods in the accuracy metric is presented in Table 4. It has been observed that the ensemble model outperforms all other classifiers. In addition, the word2vec feature model demonstrated better performance as compared to bag-of-word, TF-IDF, and word unigram. The reason behind achieving 88% accuracy with the word2vec feature is that this model employs the semantic/contextual information related to the language of posts/comments. Overall, the word unigram presented the lowest performance as compared to other feature models. On the other end, RF demonstrated the least performance as compared to the other four ML methods. In addition, it is observed that the word2vec feature model with all the classifiers demonstrated the best performance in comparison with other feature models. In contrast, we did not obtain promising results when all features are combined.

Similarly, the performance of five ML models with four feature models using precision metric is demonstrated in Table 5. Again, the performance of word2vec is better than all other features and achieved 88.28% precision with the ensemble model. The accuracy and precision measures justify that our results are consistent along both measures, and word2vec and ensemble models are the best feature and ML models. However, predictive performance with all features is not promising using the precision measure (i.e., 86.22% but 88.26% with word2vec).

The performance of four feature types is also compared using the recall evaluation metric for offensive language detection. 10-fold cross-validation and five ML methods are used for experimentation. The performance of the word2vec feature model is observed to be better than word unigram, bag-of-word, and TF-IDF feature models as shown in Table 6. The best performance (88.27% recall) is achieved with word2vec and the ensemble model. On the other hand, word unigram achieved the lowest recall with the RF model and the best recall with the ensemble model. The outperformance of the word2vec feature model with the ensemble model remains consistent along with the recall measure.

Tables 7 and 8 shows the performance of five ML models with four feature models, using the AUC and MCC measures. Here again, we can observe that the performance of the word2vec feature model is better than word n-gram, TF-IDF, and bag-of-word feature models. Moreover, the ensemble model presented better performance as compared to the other four ML methods as shown in Tables 7 and 8. Thus, along with six evaluation metrics, the outperformance of the word2vec feature model with the ensemble model is observed to be consistent. However, we did not obtain promising results when all features are combined.

The impact of proposed features is also investigated using the ROC curve as presented in Fig. 4. For experimental setup, the ensemble model is used as a classifier and 10-fold cross-validation is used for training and testing purposes. It is depicted in Fig. 4 that word2vec demonstrates the best performance and has covered more area under the curve. The performance of TF-IDF is comparable with word2vec features but slightly lower. We find symmetry in the results of ROC and other evaluation measures. Hence, the performance of features is consistent along six metrics. Thus, this proves the significance of the word2vec feature model and ensemble model for the detection of offensive language in Urdu.

Experiment 2: comparison with baseline

The latest work on offensive language detection in the Urdu language is presented by Akhter et al. (2020). In this section, we present a comparison of our model with Akhter et al. (2020). First, there is a difference in the nature and size of the dataset, as the baseline paper has created a dataset that was collected from comments in the Urdu language from different YouTube videos. These comments were manually collected from videos by the authors themselves. In our case, we have collected actual comments, shared by users, on 19 different Urdu websites including newspapers, political parties, etc. Therefore, our dataset presents originality not only in the context of comments but also in the text of the comments. Second, our dataset is much larger, starting from more than 800,000 comments, it has concluded to almost 200,000 comments after preprocessing (Table 2), whereas the baseline paper contains 2,000 comments in the Urdu language, and they have not mentioned the source of Roman Urdu. Third, Akhter et al. (2020) got the comments annotated by a panel of three persons who were all students, whereas we used the services of five language expert annotators. This has strengthened the quality of our labeling.

Regarding experiments, the baseline (Akhter et al., 2020) focused more on char and word n-gram feature extraction techniques and used several (more than 15) classifiers, whereas, in our work, we have focused on a variety of the latest feature extraction techniques and used more relevant classifiers generally used in literature for a similar task. In this context, we have performed experiments of Akhter et al. (2020) using our dataset and reported only those experiments which outperformed. Both the word n-gram-based and character n-gram-based, experiments are reported in Table 9. Our model with four feature extraction methods and ensemble as a classifier is also presented and our model has demonstrated better results as compared to the standard baseline (Akhter et al., 2020). The AUC and MCC performances are also reported. We have used six performance metrics for the comparison of results. In addition, the best score against each performance metric is also highlighted in Table 9.

It is depicted in Table 9, that our model has outperformed the baseline along with all performance metrics. The improvement is 3.55% in accuracy, 3.68% in recall, 3.60% in f1-measure, 3.67% in precision, and 2.71% in AUC. One important point in Table 9 is that the performance of the baseline technique on our dataset is less than those reported in their paper (Akhter et al., 2020). The main reason for this is the enormity of our dataset. As mentioned above, their dataset does not contain sufficient variation or originality due to how it has been generated by the authors themselves, by watching different YouTube videos. On the other hand, our dataset is made by collecting genuine texts shared by thousands of different people. This brings much more originality and variation to our dataset.

The comparison of our model with the baseline is also evaluated using the ROC curve as presented in Fig. 5. The performance of the baseline is represented by the char tri-gram in Fig. 5 because, in the baseline approach, the char tri-gram presented the best performance, as shown in Table 8 upper part. This is exactly the reason for its being selected as the baseline method. It has been shown clearly that word2vec presented better performance as compared to baseline and the other three types of features. Regarding our method, the worst performance is noted by bag-of-words, but it is still better than the baseline method. This proves the significance and effectiveness of our approach as compared to the baseline.

Experiment 3: impact of feature selection on classification performance

To reduce the complexity of feature models and to enhance the performance of the proposed detection model for offensive language, we employed the technique of feature selection (Atlam et al., 2020). A well-known feature selection method is selected to find the best subset of features. i.e., wrapper method. This method employs a search strategy (forward selection) by evaluating the possible subsets of features, using a machine learning algorithm. The evaluation of each subset is based on the quality of performance produced by the selected machine learning algorithm. The evaluation criteria may be any performance metric depending upon the nature of the problem. In our case, we have selected the SVM algorithm with an accuracy metric for the best subset selection. The reason behind choosing SVM is that it presented a significant feature subset while testing each ML model in the wrapper selection method. In the word unigram feature model, we have compared the performance of the selected subset with all Uni-gram features and did not find any improvement in the accuracy. Therefore, we did not consider the results of the word unigram.

After employing the wrapper method, we found the best subset of 67 features for TF-IDF, the subset of 72 features for bag-of-word, and the subset of 72 features for the word2vec model. Regarding performance using the accuracy metric, it is clearly shown in Table 10 that we found significant improvement in performance with the selected features for each type of feature. For example, the largest improvement is observed for the bag-of-word feature model whereas word2vec demonstrated a small improvement. An ensemble method with 10-fold cross-validation is used to generate the results.

It is clear from Table 10 that feature selection impacts all performance metrics. If we compare the result precision-wise; the bag-of-words achieved the best improvement of 3.74%, as compared to TF-IDF and word2vec with improvements of 1.81%, and .27% respectively. Similarly, the accuracy and AUC measures also improved with feature selection. Furthermore, like precision, the maximum improvement is observed with bag-of-words using accuracy and AUC metrics. These provide evidence of the usefulness of feature selection in the proposed methodology.

Experiment 4: impact of hybrid combination on classification performance

In the previous experiments, we presented the performance of the proposed features as a standalone model. In this section, we have conducted experiments to investigate the impact of various combinations of the proposed features for offensive language detection. Among four feature sets in the prior experiments, we found that word2vec has consistently outperformed the rest (with and without feature selection), whereas the performance of word unigram is the lowest among all. Ignoring the lowest one, we have made several combinations of the rest of the three feature sets. An ensemble model with 10-fold cross-validation and six evaluation metrics is used for the experimental setup.

The best performance in accuracy measure (89.23%) is achieved when we combine all three features as shown in Table 11. Similarly, this combination demonstrates the best performance on precision, recall, f1-measure, AUC, and MCC measures as well. Hence, it is the best performance obtained by our proposed model. In addition, we have observed that the second-best performance is achieved by combining bag-of-word with word2vec. The third-best performance is observed by using the combination of TF-IDF with word2vec as presented in Table 11. Thus, we can conclude that word2vec is the best feature method not only as a standalone model but also achieved the best performance when used with other features in combination. Although all evaluation metrics presented very promising values, AUC (96.78) value is very significant. Thus, the combination of the features proved its significance for offensive language detection in Urdu.

Examples of offensive and not-offensive posts

After an exhaustive evaluation of the proposed model, we have added here the predicted class labels of six randomly selected posts/comments from the test set as shown in Table 12. If we analyze the language of comments 1 and 2, we can conclude that the presence of offensive words might have made the proposed model label these posts as offensive. However, comment 3 does not contain any offensive words but it is the context of the comment that has guided the proposed model to declare it as offensive. In addition, a similar trend is also observed for not-offensive comments/posts labeled by the proposed model. The system labeled comments 4 and 6 as not offensive because there is no offensive word in both comment. However, the class label of comment 5 is decided as not-offensive using the context of the text available.

Table 11:

Performance of hybrid combinations using five evaluation measures.

Feature Combinations	Accuracy	Precision	Recall	F1-score	AUC	MCC
Bag-of-word +TF-IDF	87.82	88.42	87.9	87.67	96.05	0.856
TF-IDF +Word2vec	88.60	88.79	88.71	88.41	95.00	0.838
Bag-of-word +Word2vec	89.00	89.50	89.20	88.93	96.01	0.849
Bag-of-word +TF-IDF +Word2vec	89.23	89.94	89.49	89.04	96.78	0.895

DOI: 10.7717/peerjcs.1169/table-11

Table 12:

Examples of predicted labels.

DOI: 10.7717/peerjcs.1169/table-12