With Arabic being a grammatical gender language, speakers use gendered pronouns, nouns, verbs, and adjectives to refer to themselves and others. The word for doctor in Arabic is either tabib (male doctor) or tabiba (female doctor), so if a question has the expression ana tabib, this indicates the questioner is male. We also use the answers for the same purpose. kama ta’alm means as you (male) know, while the female version is kama ta’lamin. We use these linguistic cues to assign gender to questions. We have a four-step approach as follows:

After applying the four steps, we found 40458 with uncontested gender information out of 172,000 fatwas. The manual checks found no incorrectly assigned fatwas.

The resulting dataset comprises 40458 questions with 17221 asked by women and 23237 asked by men. Questions vary considerably in length with the average number of words per question at 116.83, with a standard deviation of 120.25, a median of 80, a minimum of 3, and a maximum of 3019. There are differences between the lengths of questions by men and women. For men, the average is 109.5 words (std = 117, min = 3, max = 3019, median = 74) while for women the average is 126.79 (std = 123.9, min = 4, max = 2109, median = 89). The dataset is distributed in a JSON streaming file with the keys: question, answer, date indicating the date of publication, classification, which is a coarse classification of the question provided by the website and views, which is the number of times the fatwa has been viewed. The dataset (40458 questions) has been randomly divided into a training set (85%, 34389 questions) and a test set (15%, 6069 questions). 10% of the training set has also been dedicated as a development set. This division is the same across all the experiments below. All the dataset metadata is available, directly or indirectly in the html. In Figure 1, we have a screenshot of a recent fatwa with numbers marking the different pieces of the annotation:

A white space-delimited unit in Arabic is usually made up of zero or more prefixes, a lexical item, and zero or more suffixes. Common prefixes are conjunctions, prepositions, and the definite article while common suffixes are possessive and object pronouns. For example, the orthographic unit fsykfykhm, depicted in Figure 3, is made up of the conjunction f, the future particle s, the verb ykfy, the direct object k and the indirect object hm. The verb itself is made up of the prefix n and the verb stem kfy. In most of what we do in this paper, we are mostly concerned with the stem.

For stemming to happen, the text is first run through morphological segmentation, which sets segment boundaries within words. The segments are then passed through a part of the speech tagger that assigns grammatical tags (e.g. NOUN, VERB, ADJECTIVE) to these segments. The stem is the main lexical unit in the word and must have a lexical tag (NOUN, VERB, ADJECTIVE). All other tags are discarded in the topic modeling experiments while in lexical experiments they are retained since they are useful for style differentiation. Stemming is achieved through the use of ArabicSOS, which is specialized in segmentation, orthographic standardization, and stemming of classical and religious Arabic (Mohamed and Sayyed). The stemming effect on the dataset is dramatic as the number of words in the questions is 4729279 with 168827 unique words. The number of segments is 7868382 with 48553 unique segments. The unique segments are hardly 29% of the original unique words.

Topic modelling is a way of summarising text documents into collections of thematically related words called topics. Due to the nature of Arabic, there is a high type-token ratio as there are too many unique words, and running topic modelling with this is not very useful. For this reason, we use the stems as input to the topic modelling software. We use Mallet (McCallum) for running the topic modelling and we are interested in two Mallet outputs: (1) the keys of the topics, or the lexical items constituting the topic, and (2) the probabilities associated with the topics, especially the probability of each topic in each document as we use these probabilities as input to the machine learning classifiers below. For consistency in topic modelling, we use the same 50 topics for all the experiments below.

The purpose of regression is to predict, for each question, how popular the question may be. Just like with classification, we use both the textual content and the topics for regression. In both regression and classification, we use a variety of algorithms that will be detailed below. For both regression and classification, we use 80% of the data for training and 20% for testing. We mainly use the scikit-learn (Pedregosa et al.; Mohamed and Sarwar) library, which has a wide range of algorithms.

What, if any, are the differences between the questions posed by women and those asked by men?

In order to find out what the differences may be between the questions asked by women and those asked by men, we use topic modelling. We then use Odds Ratios (see Tables 3 and 4) to find out which topics are more correlated with men and which are more characteristic of women. Table 1 lists the 50 topics produced by Mallet ordered by their odds ratios. Since female-centric questions were assigned the zero class, odds ratios less than one are more associated with women while odds ratios larger than one are more characteristic of male-centric questions.

We can see from the table that there are clear differences between the questions asked by women and those raised by men. The top concern for women in this dataset is family matters (children, marriage, divorce). We can also see religious ritual questions concerning menstruation and fasting as menstruating women do not have to observe the obligatory fasting in the lunar Islamic month of Ramadan. The current study focuses on a convenience sample of Muslim men and women who are mostly in the Middle East, and the results may thus be hard to generalize. The interest in the family may echo women’s interests in other cultures as well. A study that examined a corpus of happiness moments found out that women and men have different sources of well-being (Mohamed and Mostafa), as men’s top source of happiness was related to games and sports while women’s was more related to family and shopping.

While differences between Men and Women may not be news in general, the details of these differences within the religious domain are of interest. For example, Men seem to be interested in totally different things, such as ensuring that the job or work they have is conformant to the rules of Islam, working and residing in other countries, especially Western countries, and obtaining their residences and citizenship, marriage and divorce consequences, the differences among scholars of Islam, life, death, and the hereafter, the relationships between Muslims and non-Muslims, banking and usury, especially concerning getting a bank loan and whether this is permissible in Islam, trade and transactions and rules on how to make profit and the permissibility of commissions, and questions about praying in mosques vs. praying at home. For example, a representative fatwa for topic 48, the second most male topic reads:

While a representative fatwa for topic 4, the most female topic reads:

The use of topic models also gives us the chance to examine which topics go together. Figure 5 shows the relationships between the topics. To create this network, we only considered the most probable topic pairs. To illustrate, let’s examine topic 4, the most common topic in the questions asked by women. When we have topic 4 as the most probable theme in a certain document, topics 13 and 26 are the most probable second topics in that document, so we connect them together. For topic 4, this results in the graph shown in Figure 4. What this indicates is that if a question involves matters of problems with their husband, family, children, and divorce, then it is also very likely to mention that the husband hits the questioner, insults her in front of other people, is mad at her and treats her badly (topic 13), and that this may come in the context of a discussion about visiting family (Topic 26). Topic 4 is also very likely to be invoked in the context of topic 44 (divorce anger return home dispute), topic 38 (money salary haram amount expenses help), topic 49 (tradition problem talk people hope to return), and topic 17 (life problem big family tries to leave).

Can we automatically classify the questions as either male or female?

In our specific case, we want to predict the gender of the author of the question even though we may use the answer as a means to this prediction. If the answer is addressed to a female, based on the morphology, then it is certain that the question was asked by a female. If the addressee is male then the question was asked by a male. In the cases where the morphology does not help identify the gender of the questioner, we do not use that specific Fatwa in training the machine learning classifier.

We run several experiments to predict gender. In all cases, we use the Term Frequency Inverse Document Frequency (TF-IDF) for feature extraction and an n-gram range of (1, 2), which means we use both individual words and two consecutive words as features. we use three algorithms in this task: Logistic Regression, Random Forests, and Support Vector Machines. As far as the input text is concerned, we vary the input to be either the question alone, the answer alone, or a combination of the answer and the question.

Table 2 lists the results of these experiments. Using both the questions and answers as input with the Support Vector Machines algorithms yields the best results with precision, recall, and an F1 score of 0.98, which is a very high number, and means that we can predict the gender of the one who asked the question with 98% score. Using topic probabilities in gender classification did not yield good results, compared to lexical input, as the best result was an F1 score of 0.68 using Logistic Regression.

While logistic regression is not the best-performing prediction algorithm, one major advantage of its use is its high interpretability. One can use the coefficients of the features (i.e. the lexical items) to see which words are more likely to be used by males and which are more often used by females. This also applies to the words and phrases more commonly used by the muftis when the questioner is male or female. This also helps with the possibility of generalization. We have extracted 40,000 out of 170,000 fatwas based on morphological clues. If these morphological clues distinguish male from female fatwas, then the possibility of an accurate classifier that generalizes beyond this small corpus is limited. If, on the other hand, the features are not limited to the morphological cues then we can use the classifier to predict much more than the small dataset. Examining the top 100 male lexical features and their equivalent female ones shows that the classifier is capable of generalization. The top 10 male features are not morphological.

Can we predict the popularity of a fatwa based on its content? To answer this question, we perform text regression where the predictor variables are the lexical items of the questions, and the response variable is the number of views each question receives.

To perform the regression, we convert the text into vectors. We do so through TFIDF as explained above. When we compute the number of views, we do so based on the number of views per day. Since the fatwas appeared online on different dates, it may not be fair to compare the raw views. This is the reason we divide the number of views by the number of days a fatwa is online, so we may obtain daily views. The daily views, rather than the raw ones, are what we want to predict. As shown in Figure 2, the fatwa dates range between 1999 and 2015, but the views were recorded in February 2016 and July 2020. Since dates of publication are available for each fatwa, we subtract the publication date from 8/7/2020, the days the fatwa was downloaded, to obtain the number of days the fatwa was online. We then divide the number of views by the number of days to obtain the views per day. The views per day is the number we predict using the various regression models. We use different algorithms and compare the results to see which one best predicts the daily views per question. We also examine the feature ranking for features that affect the regression model. In terms of evaluation, we evaluate the regression model based on two things: (1) the R2 value and (2) how good it is at predicting the views in a test set based on the Mean Absolute Error (MAE). The R2 is a measure of goodness-of-fit and is usually interpreted as how much variation is explained by the independent variables. An R2 of 0.6, for example, indicates that 60% of the variation in the dependent variable can be explained by the independent variables. The R2 is known to be sensitive to the number of independent variables as it increases proportionally with the increase in the number of independent variables. This is a problem in text regression as the number of independent variables is very large, which may lead to inflated R2 values. For this reason, we also use prediction to evaluate the regression model. The MAE is a measure of the average distance between the actual values and the values predicted by the model and can thus be useful in comparing various models.

The results of regression, as shown in Table 5 indicate that when we use a combination of word unigrams and bigrams, both linear regression and Random Forests regression do a good job predicting views per day given the question as input. Linear regression has an R2 of 0.99953, which is near perfect but may also indicate over-fitting for the training data. The mean absolute error is 5.62 on the test set. This is a good MAE given that the standard deviation of the test set values is 9.55. Random Forests do an even better job at prediction, probably due to their non-over-fitting, with an MAE of 3.49. This indicates that predicting the views per day based on the textual content is feasible. The experiments also show that segmentation does not help as every whole-word experiment yields better results than its segmented counterpart. We attribute this to the availability of data. With our large data set, there is no need for segmentation, whose main purpose is to combat data sparseness.

The Random Forests algorithm is superior to the linear regression one in prediction. One other nice facet of RF is that it also produces a list of the top features used in regression. Perhaps one should examine what features in the input are responsible for these views. For this, we use the most important features (lexical n-grams) as produced by the RF algorithm. An examination of the top 100 features in the unigram +bigram model reveals that the most important concepts that trigger views are related to sex, cleanliness, and prayer. For example, the top 10 grams can be translated into: through the condom, is foreplay permissible, caressing the, anal, condom, inserting part of, intercourse, prayer, at the right time, masturbation, and orgasm.

Perhaps a more effective way for finding out popularity triggers is to find what themes are more viewed. For this purpose, we use the probabilities of the topics produced by topic modelling as features and the views per day as the dependent variable, to rank the topics in terms of their importance to the Random Forest regression algorithm. This ranking shows how much each topic contributes to the number of views per day. Table 6 lists the five most important topics, and we can see that questions about marriage and engagement top the list. A common theme here is the problem of the family refusing the prospective fiance(e), and the question is usually what one should do in such a case. This is followed by the question of (premarital and extra-marital) sexual activity and how one should repent. We then have topics on cleanliness and prayer, fasting the month of Ramadan, and what one should do if they miss days of required fasting, and then we have financial questions most of which are about whether some financial transaction is halal (permissible in Islam) or haram (impermissible according to Islamic law). The prominence of sexual topics may be related to the fact that Islam is very restrictive in terms of premarital sex (Adamczyk and Hayes).