Can a Machine Learn to Classify Meccan and Medinan Surahs?

A basic concept in Quranic studies is the difference between its meccan surahs/chapters (revealed before the Prophet Muhammad migrated from trading city of Mecca to oasis town of Medina) and its medinan chapters (revealed after the migration).

Knowing where a chapter was revealed helps scholars derive meaning from the text, and so a great deal of effort has been expended to finding historical narrations that mention when a particular verse or chapter was revealed. However, the location where a chapter was revealed is also manifested in the linguistic characteristics of the passage. For example, meccan chapters are characterized by poetic meter, a rhetorical urgency, and an emphasis on recognizing the Oneness of God, while medinan chapters are more lengthy prose dedicated to explaining religious rituals.

This made me think that it might be possible to build a binary classifier that could use basic techniques from natural language processing to classify chapters as medinan or meccan. Here, I outline my methodology and share my results.

Methodology

Downloading the Dataset. The first step is to download all of the text of the Quran. The standard text of the Quran can be downloaded from:

http://tanzil.net/docs/download

The version without vowel marks and with verse numbers will work best for our work. We also need to know which surahs are meccan and which are medinan. Find this information at:

http://tanzil.net/docs/revelation_order

Alternatively, a link to a github repository with the .csv files (and all source code) is included at the end of this post.

Word Vectors Once we have the text of the Quran, we need to convert the text to a standardized representation that can be fed to machine learning algorithms. Usually, these representations are in the form of numeric arrays, matrices, or tensors.

There are a variety of ways to do this, but the simplest may be the bag-of-words technique, which represents a ‘text’ (in this case, we’ll use each verse as a text) as a vector, with each dimension representing the occurrence of a particular word.

As an example, let’s say the words in our body of text are:

إنا – الإنسان – لفي – خسر – أعطينا – الكوثر

Then, we could represent these two verses as follows:

Verse Bag of Words Vector
إنا الانسن لفى خسر [1, 1, 1, 1, 0, 0]
إنا اعطينك الكوثر [1, 0, 0, 0, 1, 1]

If a word repeats more than once, the vector reflects that multiplicity, so a vector can have entries more than 1 as well.

Training and Validation

Before feeding the word vectors into a classifier, we need to partition our dataset into a training and validation set. To ensure that the validation set and training set are independent, the partitioning is done on a chapter level, not a verse level. This is because some chapter include repeating refrains (e.g. Surah Al-Rahman), it is unfair to include those verses in both the validation and training sets.

The partitioning I used was: Training: 40%, Validation: 60% (I chose a larger validation set to increase the meaningfulness of results, as will be discussed later). Because the partitioning was done on a chapter level, the number of verses in the training and validation set was not exactly 40/60.

Logistic Regression

The classifier I ended up choosing was logistic regression. Logistic regression works similarly to linear regression, as it assigns linear weights to the presence of each feature in our vector. But then it introduces a non-linearity – by passing the result through the logistic function

This provides a probability between 0 and 1 that verse belongs to a particular class – in our case, the classes are meccan or medinan. We train the weights on our training set, and check how well it performs on the test set.

Results

So how did we do? The following table represents the training and validation accuracies on a typical run:

Round Accuracy
Training 98.68%
Validation 86.26%

The training accuracy is very high because we are over-fitting on our datasets. This not surprising because the number of parameters in our model – or the number of unique words in the Quran – is 14,870, far more than the number of verses in our training set.

However, we do pretty decently on our validation set as well. The accuracy, 86%, needs to be taken within context:

  • The dataset is unbalanced – there are many more meccan chapters than medinan chapters, so if our classifier was just classifying every verse as meccan, it would get an accuracy of about 74%. Depending on the specific partitioning of the training and validation, we have an accuracy around 10 percentage points higher than that.
  • This is on a verse-by-verse level, under the assumption that every verse in a meccan surah is meccan, and same for medinan. This assumption is definitely not true, which affects both our training and evaluation of performance.

A better indicator would be performance on a surah level. If we use a simply majority voting system (using all the verses in a chapter to “vote” for whether the surah is meccan or medinan), we get the following results:

Round Accuracy
Overall surah-level 94.73%

I’ll confess, I was a bit disappointed when I saw that this wasn’t closer to 100%. Which surahs are misclassified? It turns out that there are 6 surahs that are misclassified in the validation set:

Misclassified Surah Predicted type Confidence (votes)
Surah Ra’d (13) meccan 81%
Surah Hajj (22) meccan 63%
Surah Muhammad (47) meccan 53%
Surah Rahman (55) meccan 100%
Surah Insaan (76) meccan 97%
Surah Zilzaal (99) meccan 100%

When I took a closer look at these chapters, I was amazed to see that there is actually scholarly disagreement about all 6 of these chapters! Based solely on word usage, our classifier suggests that a few of these, such as Surahs Rahman, Insaan, and Zilzaal might fall in the meccan camp!

Finally, I was curious to see which Arabic words are the “most meccan” and the “most medinan.” We can analyze this by looking at the weights that the model learns for each word. The top 10 meccan and medinan words are:

Top Meccan Words Top Medinan Words
بعهدكم تقتلني
وتقسطوا والله
العالمون أكلها
أولاهما وإذ
لتحصنكم وبكفرهم
العمى مبصرة
نور زلزلة
لمستقر بآية
لنفد سخرناها
تنزيل وجهرا

Perhaps those who have studied the Quran more can enlighten me, but these words don’t seem to follow any significant trends.

I did notice that the words in the meccan column are mostly those words that are rare, but make an appearance in at least once meccan surah. For example, لنفد is found only in Surah Al-Kahf. The words in the medinan column are sometimes also hapaxes, but in other cases, they are common words in the Quran, which are overweighted in the medinan verses, which tend to be longer than meccan verses on average. So while والله would appear in both kinds of surahs, it might appear multiple times within a medinan verse. A further confounding factor is that our simple “bag of words” technique is unable to recognize any degree of similarity between related words – والله is a separate word than الله internally – making it more difficult to build complex insights.

With a deeper look into the weights, one might be able to identify trends as to why certain words are more “meccan” and why others are more “medinan,” but the most strongly weighted words seem to result of technicalities associated with bag-of-word techniques. (Perhaps normalizing vectors before feeding them into logistic regression would lead to a more interpretable model?).

Conclusion

The basic conclusion is that we can design an extremely simple algorithm to classify meccan and medinan surahs, based on a small set of training examples! While its not clear the internal model that the algorithm builds is particularly sophisticated or enlightening, we see that it performs quite well – only incorrectly classifying 6 surahs, all of which are the subject of scholarly debate anyway!


All the source code and data files for this project can be found and cloned from this github repository.

Written on January 16, 2017