1 Introduction
The Quran, a religious text revealed to the Prophet Muhammad (Peace Be Upon Him) over 14 centuries ago, holds immense significance. Muslims widely read it for its spiritual importance and comprehensive guidance in addressing various challenges across social, economic, and religious aspects of life. Religion, embraced by diverse nations speaking different languages, translates its religious texts into numerous languages to guide its followers. However, translating a spiritual text can lead to varying interpretations due to the background and understanding of the translator.
Scholars delve into Quranic verses based on their areas of interest. For instance, some scholars focus on “mercy” in the Quran to identify verses, passages, and interpretations that emphasize mercy. Manually identifying topics and conducting in-depth studies from different Quranic translations is time-consuming because the meanings and ideas of different translators overlap from verse to verse and from chapter to chapter [Al Ghamdi and Khan
2022]. Hence, an automatic method is needed to identify topics from multiple translations of religious texts. One solution to this problem is topic modeling, which extracts hidden topics from an extensive collection of documents [Alhawarat
2015]. It allows scholars to gain valuable insights for focused research [Siddiqui et al.
2013]. Topic modeling has been applied to Quranic translations in various languages such as English [Al Ghamdi and Khan
2022], Arabic [Alshammeri et al.
2021b], and Indonesian [Rolliawati et al.
2020]. Still, such techniques are not available for Urdu translations of religious texts.
Urdu, a language rich in grammar but lacking in resources, poses challenges due to its numerous word derivations and inflections [Khalid et al.
2017]. Unlike English, Urdu is written from right to left and follows the Nastaliq writing style, similar to Arabic [Alhawarat and Hegazi
2018; Al Qudah et al.
2022]. Therefore, many essential resources and accurate text processing toolkits developed for other languages cannot be directly applied to Urdu [Daud et al.
2017]. Limited research on Urdu topic modeling has mainly focused on classical approaches [Shakeel et al.
2018; Rehman et al.
2018; Amin et al.
2020]. Topic modeling on religious texts is more challenging than on general texts because a single topic may repeat in different places in the Quran with varying details and contexts. Additionally, one verse may contain multiple topics, making it even more challenging due to the variations in Quranic translations. Although traditional techniques such as Latent Semantic Analysis [Landauer et al.
1998], Probabilistic Latent Semantic Analysis [Hofmann
1999],
Nonnegative Matrix Factorization (NMF) [Lee and Seung
2000], and
Latent Dirichlet Allocation (LDA) [Blei et al.
2003] are widely used and have provided valuable insights, they often struggle to capture contextual nuances and semantic representations present in Urdu translations of the Quranic text [Vayansky and Kumar
2020]. In recent years,
Bidirectional Transformers for Language Understanding (BERT) [Devlin et al.
2019] has shown promise for context-aware language understanding, successfully applied in English and Arabic. It highlights the need for further research to leverage the potential of such models in Urdu translations of the Quran.
In this work, by leveraging BERT contextual understanding, we have proposed a topic modeling technique to analyze themes, concepts, and underlying meanings in the Urdu translations of the Quran. Our contributions lie in exploring the adaptability of BERT-based models to the intricacies of the Urdu language and Quranic text and in capturing the nuanced semantics of Urdu Quranic translations, thus offering a valuable resource for scholars, researchers, and practitioners of religious studies. Through rigorous experimentation, we demonstrate the potential of these models, therefore advancing the state of the art in text analysis for religious texts in Urdu while addressing the existing research gap.
The contributions of this work are outlined as follows:
—
We develop the
Quran Urdu Topic Modeling Corpus (Quran-UTM), which comprises Urdu translations of the Quran, including eight esteemed translations by widely recognized translators. To focus on the actual content for topic modeling, the irrelevant information such as metadata, punctuation, and diacritics are removed from each Urdu translation. Corpus and source code are publicly available for the research community.
1—
We have developed a framework for context-aware topic modeling using BERT.
—
We have assessed the performance of our BERT-based model against traditional methods, such as LDA and NMF. The results show that our BERT-based approach outperforms these conventional methods, achieving significantly higher coherence and diversity scores.
This approach has the potential to bridge the existing gap in Quranic topic modeling for Urdu translations and unlock new insights into the Quran’s profound message in the Urdu language.
The rest of the article is organized as follows. Section
2 presents a comparative review of existing work in topic modeling. Section
3 presents the Quran-UTM dataset. Section
4 presents the proposed methodology, and Section
5 discusses the experimental setup and presents the obtained results. Section
6 provides a discussion and analysis of the results, and finally, Section
7 concludes the study and offers suggestions for future research directions.
2 Related Work
Several studies have been done on topic modeling of Quranic text and its translations in recent years. While most of these studies have focused on the original Arabic script [Abuzayed and Al-Khalifa
2021] and English translations [Alshammeri et al.
2021b] of the Holy Quran, some have also explored other domains. This section provides an overview of previous works that deal with traditional topic modeling techniques for Quranic text and its translations and the application of context-aware advanced methods.
2.1 Classical Techniques for Topic Modeling
Classical techniques for topic modeling have been employed for text in multiple languages, including English, Arabic, and Urdu.
Rahman et al. [
2018] aim to classify topics in each verse of Surah-Al-Baqarah from the Quran using machine learning techniques, particularly employing
Support Vector Machine (SVM) as the primary algorithm. A comparative analysis with other classification algorithms is conducted. The research contributes to developing a computational environment for text mining the Quran and enabling efficient topic identification. However, there is a need for further investigation to explore additional methods or approaches that can enhance the accuracy and comprehensiveness of topic classification in Islamic literature.
Siddiqui et al. [
2015] apply a probabilistic topic modeling algorithm, LDA, to discover the thematic structure of the Quran. By considering each Surah as a document, the paper aims to automatically identify the hidden thematic structure of the Quran, which contains vast information addressing various aspects of human life. Using the Arabic Quran as the corpus, the study successfully identifies significant themes in the Surahs and determines the most important terms associated with these themes.
Seifollahi et al. [
2021] introduce a two-stage algorithm for topic modeling, addressing the limitations of traditional methods. While LDA relies on simplistic bag-of-words assumptions, this innovative approach leverages word embeddings and co-occurrence patterns. The first stage determines topic–word distributions by soft-clustering embedded n-grams from documents. In the second stage, it computes document–topic distributions by sampling topics from the topic–word distributions. This method capitalizes on word embeddings, providing a more accurate representation. Nonetheless, the article does not explicitly address potential constraints or difficulties linked to the proposed approach, leaving an opportunity for further investigation in future research.
Alshammeri et al. [
2021a] employ the
Natural Language Processing (NLP) method Doc2vec to detect semantic-based similarity between verses of the Quran. By mapping Arabic Quranic verses to numerical vectors encoding their semantic properties, the model achieves 76% accuracy in similarity detection compared to annotated textual similarity datasets. These findings provide a foundation for further research in Quranic semantic analysis, although improvements are needed to enhance the model’s performance and robustness.
2.2 Context-aware Techniques for Topic Modeling
In recent years, researchers have focused on using context-aware techniques for topic modeling for text in various languages.
Abuzayed and Al-Khalifa [
2021] conduct an experimental study on BERT for Arabic Topic Modeling, specifically focusing on the effectiveness of the BERTopic technique. The study aims to explore the application of BERTopic using different Pre-trained Arabic Language Models as embeddings. By evaluating the results against traditional techniques like LDA and NMF, the study highlights the superior performance of BERTopic in uncovering latent topics in Arabic text. The evaluation metric was the
Normalized Pointwise Mutual Information (NPMI) measure. This research contributes to the literature by emphasizing the potential of BERT-based techniques for Arabic topic modeling and addressing a research gap in this area.
Alhaj et al. [
2022] propose a deep-learning-based approach using BERTopic to improve the classification of cognitive distortions in Arabic content on Twitter. By enriching text representation with latent topics derived from BERTopic, the study achieves improved performance compared to baseline models. However, more research is needed to understand the specific impact of BERTopic on cognitive distortion classification in Arabic on social media platforms. Further investigation is required to explore this aspect and provide deeper insights.
Alsaleh et al. [
2021] introduce AraBERT, a powerful language model based on BERT that has demonstrated remarkable performance in various Natural Language Processing tasks. In this study, the authors leverage AraBERT to classify pairs of Quranic verses from the QurSim dataset as semantically related. The authors pre-processed the dataset and created three subsets for comparison. AraBERT, AraBERTv0.2, and AraBERTv2 are evaluated to determine the optimal version for our datasets. AraBERTv0.2 achieves an impressive accuracy score of 92% on a dataset containing labels “2” and “–1,” the latter generated outside the QurSim dataset. While this paper successfully explores Quranic verse semantic relatedness using AraBERT, there is still room for further research and improvement to better understand the contextual nuances within the Quranic text.
Alshammeri et al. [
2021c] introduce a novel approach for detecting semantic similarity in the Quran using a Siamese transformer-based architecture. The study focuses on the significance of semantic similarity detection in various natural language comprehension tasks and its relevance to NLP applications. Leveraging Arabic pre-trained contextual representations, the proposed model generates verse embeddings that capture semantically meaningful information. The twin transformer networks are fine-tuned on a Quranic semantic similarity dataset, resulting in improved performance compared to previous studies. However, the article does not explicitly address any potential limitations or challenges associated with the proposed approach, leaving an opportunity for further investigation in future research.
The article of [Yang et al.
2023] introduces sDTM, a novel supervised topic modeling approach that enhances traditional methods like LDA by incorporating auxiliary data such as star ratings or post categories. Combining a neural variational autoencoder and a recurrent neural network, sDTM improves empirical estimation and predictive performance. While making valuable contributions to the field, the article could benefit from further investigation into scalability and practical limitations, ensuring its applicability to large datasets and real-world text analytics tasks. In their study, George and Sumathy [
2023] incorporated LDA with clustering, BERT, and dimensionality reduction techniques (PCA, t-SNE, UMAP) to enhance topic modeling. The experiments were conducting using the open-source, cross-platform
Integrated Development Environment (IDE) Spyder. The machine learning experiments were performed using Scikit-learn and NLTK tools for analysis. While the results demonstrated improved topic extraction on benchmark datasets, there is a need for further quantitative analysis of topic quality enhancement, as well as considerations for scalability and real-world deployment challenges.
Table
1 provides a quick comparison of existing research categorized by language (English, Arabic, Indonesian, and Urdu), arranged in reverse chronological order within each language group based on their publication year. Notable observations from the existing studies include:
—
Limited Research on Al-Quran Urdu Translation. As shown in Table
1, there are few studies on Urdu topic modeling between 2018 and 2023, and these studies do not cover Quranic text, which indicates the research community’s limited focus on this specific NLP task. Conversely, topic modeling on religious texts has been conducted in various languages such as English, Arabic, and Indonesian, highlighting the need for more comprehensive investigations into Al-Quran Urdu translation.
—
Scarcity of Context-aware Topic Modeling Techniques for Urdu. Table
1 reveals that many studies in English and Arabic have applied context-aware techniques, especially transformer-based language models like BERT and its variations. Most of these studies employing context-aware topic modeling techniques have been published between 2021 and 2024, demonstrating the rapid adoption of these models in the research community. In contrast, only one recent study in Urdu utilized static word2vec embeddings [Mustafa et al.
2023]. Due to variations in Quranic translation, identifying topics from different Quranic translations is more challenging. Static embeddings assign fixed vectors to words regardless of context, lacking adaptability, while transformer-based contextual embeddings dynamically capture nuanced meanings by contextualizing words within various Quranic Urdu translations. Consequently, there is a need to evaluate the effectiveness of context-aware topic modeling techniques in the Urdu translation of the Quran, which is the focus of this research.
In conclusion, context-aware techniques have exhibited remarkable effectiveness in topic modeling for religious texts, particularly in English and Arabic, when applied to the Holy Quran. However, utilizing these techniques for topic modeling in Urdu within the religious context is an under-explored research area. Our study aims to bridge this gap by leveraging BERT-based approaches for text analysis and topic modeling for Urdu translations of the Quran. Our contributions lie in not only exploring the adaptability of BERT-based models to the intricacies of the Urdu language and Quranic text but also capturing the nuanced semantics of Urdu Quranic translations, thus offering a valuable resource for scholars, researchers, and practitioners in the field of religious studies. Through rigorous experimentation, we demonstrate the potential of these models, thus advancing the state-of-the-art text analysis for religious texts in Urdu while addressing this pressing research gap.
4 Proposed Methodology
This section presents our proposed BERTopic model [Grootendorst
2022] for the Quranic Urdu translations topic modeling. To capture the contextual nuances of the Quran in Urdu, we utilize BERT embeddings, which convert the Urdu translations into semantic representations in vectors. Afterward, we reduce the dimensionality of the vector representation through
Uniform Manifold Approximation and Projection (UMAP), followed by clustering to group verses with similar semantics. We use descriptive analysis and evaluation metrics to infer and assess the discovered topics. Figure
3 provides a workflow of our proposed methodology. The details of each step are provided in subsequent sections.
4.1 BERT Embeddings
First, we transform our documents into numerical representations using BERT embedding, which differentiates the different meanings of the same word and various words with the same meaning based on their context within a sentence.
For example, as illustrated in Figure
4, the word
in the Surah Taha represents light and guidance. It symbolizes the divine illumination and the hope for spiritual insight that Prophet Musa sought when he approached the fire on Mount Sinai. Conversely,
in Surah Al-Baqrah verse number 24 denotes Hellfire, representing the concept of punishment and suffering in the afterlife. These two contexts show how the same word,
, diverges into vastly different topics in topic modeling. One context is associated with light, guidance, and knowledge, while the other pertains to punishment, retribution, and the concept of Hell. It illustrates the richness of language in the Quran, where a single term can encapsulate multiple layers of meaning depending on its usage.
Different translators choose different words to translate the original text based on their understanding. For example, in Figure
5, the words
and
in these translations serve the same purpose of conveying a successful outcome for those who follow divine guidance. These words belong to the same topic. Traditional techniques failed to capture the contextual nuances and semantic representations in different Urdu translations of the Quranic text. BERT achieves this through bidirectional analysis, considering preceding and following words to generate context-specific embeddings.
To apply BERT embedding, we utilized the pre-trained sentence transformer model “paraphrase-multilingual-MiniLM-L12-v2” from Hugging Face [Reimers and Gurevych
2019]. Hugging Face is a popular library and platform that provides a wide range of pre-trained models for natural language processing tasks. Sentence transformers, a type of deep learning model, have gained significant attention for their ability to capture semantic representations of sentences and paragraphs. The sentence transformer model we employed, specifically the “paraphrase-multilingual-MiniLM-L12-v2,” is designed to map sentences and paragraphs to a 384-dimensional dense vector space, enabling clustering or semantic search tasks. By leveraging transformer architectures, sentence transformers encode contextual information and the semantic meaning of the text into these dense vector representations.
In our case, the sentence transformer model was employed to generate BERT embeddings of the proposed Quran-UTM dataset. These embeddings serve as high-dimensional representations of the text, capturing the intricate relationships between words and phrases. By leveraging the contextual information encoded in the embeddings, we can understand the finer details and subtle aspects of the Quranic text in Urdu, uncovering its hidden meanings and more profound significance.
4.2 Dimensionality Reduction
When using BERT embeddings, each word or token of the Quran-UTM dataset is represented by a high-dimensional vector. This high dimensionality can pose challenges regarding computational complexity, memory requirements, and the interpretability of the embeddings. A dimensionality reduction technique is employed to handle the high dimensionality of BERT embeddings. We utilized the UMAP model for dimensionality reduction [McInnes et al.
2018]. There are traditional approaches to reduce the dimensionality of document embedding, such as
t-Distributed Stochastic Neighbor Embedding (t-SNE), which emphasizes preserving local similarities [Van der Maaten and Hinton
2008], and
Principal Component Analysis (PCA), which is a linear dimensionality reduction technique that focuses on the global structure of the data by projecting it into a lower-dimensional space [Pearson
1901]. In contrast, UMAP offers a balance between preserving global and local structure in data, making it suitable for effectively visualizing and analyzing complex datasets [McInnes et al.
2018]. Additionally, research has shown that reducing high-dimensional embedding with UMAP can enhance the performance of clustering algorithms like K-means and
Hierarchical Density-Based Spatial Clustering of Applications with Noise (HDBSCAN) in terms of accuracy and processing time [Allaoui et al.
2020]. UMAP uses a cross-entropy-based loss function to optimize the low-dimensional representation. The loss function is defined in Equation (
1):
where:
—
\(i, j\) represent pairs of data points.
—
\(d_{ij}\) is the distance between data points i and j in the high-dimensional space.
—
\(\sigma _{ij}\) is the similarity between data points i and j in the low-dimensional space (after applying a Gaussian kernel).
—
\(w_{ij}\) is a weight that adjusts the importance of each pair of data points based on their distance in the nearest neighbor graph.
UMAP works by mapping the high-dimensional space of BERT embeddings onto a lower-dimensional space, typically two or three dimensions, which can be easily visualized. The mapping process considers the underlying manifold structure of the data, capturing the essential similarities and differences between the verses. By reducing the dimensionality, we can simplify the data representation, making it more manageable for further analysis and interpretation.
4.3 Clustering
Clustering plays a crucial role in the BERTopic model as it enables grouping similar Quran verses based on their semantic similarities. This step is essential for organizing the vast corpus of Quran-UTM into meaningful clusters or groups, each representing a potential topic. A suitable algorithm is employed for clustering, such as K-means or HDBSCAN. By default, the BERTopic framework uses HDBSCAN clustering [McInnes et al.
2017]. HDBSCAN is a density-based clustering algorithm that builds upon the traditional DBSCAN algorithm by incorporating a hierarchical approach. It has the ability to handle clusters of varying densities and identify outliers. Sometimes, HDBSCAN produces numerous outliers, potentially leading to excessive outlier clusters, making it difficult to identify meaningful clusters accurately. So, in this specific study, we choose the K-means algorithm because it prevents the generation of excessive outliers. It is a partition-based clustering algorithm that iteratively assigns data points to the nearest cluster centroid to minimize the sum of squared distances within each cluster.
By applying K-means to the reduced-dimensional BERT embeddings, we discover distinct clusters of Quran verses that exhibit similar semantic characteristics. These clusters represent topics or themes that emerge from the Quran Urdu translations. By grouping verses with shared semantic similarities, we gain valuable insights into the underlying structure and content of the text, facilitating the exploration and analysis of the translations. Clustering helps to organize the Quran Urdu translations into coherent groups, making it easier to identify patterns, trends, and topics within the text. It provides a framework for understanding the relationships between different verses and unveiling the underlying themes present in the translations.
4.4 Tokenizer
To ensure modularity in BERTopic’s algorithm, a flexible topic representation technique is essential, one that avoids assumptions about cluster structures. We employ a bag-of-words approach, where all documents within a cluster are combined into a single document to count word frequencies, generating a cluster-level representation. This process uses the default tokenizer from CountVectorizer in the sklearn library,
5 which breaks down the text into individual words or tokens and handles tasks like stop word removal.
Stopwords, being less meaningful and the most frequent words in any text, connect the significant parts of a sentence and build context. However, our approach uses sentence transformers that extract contextual information from sentences, making the removal of stopwords in the initial stages unnecessary.
Eliminating stopwords during the post-processing phase was crucial for enhancing the quality and relevance of the identified topics by removing less informative words and focusing on more meaningful content in the Quran-UTM dataset. In our proposed framework, we utilized a freely available, pre-existing stopwords list [Shafi et al.
2023].
4.5 Topic Inference
After tokenizing and removing the stopwords within a cluster, the next step is to analyze the verses within each cluster to identify the most representative words and themes. Topic inference is the process of identifying and extracting meaningful words from clusters. It involves analyzing the content and patterns within the data to uncover underlying themes, concepts, or subjects prevalent across the clusters of verses.
Descriptive analysis techniques are applied to the verses within each cluster during topic inference. It involves calculating the importance or significance of words within a cluster using Class-based Term Frequency-Inverse Document Frequency (C-TF-IDF).
The C-TF-IDF weight for a term
t in a document
d belonging to a cluster
c is given by Equation (
2):
By examining the C-TF-IDF scores of words within a cluster, we can identify the most characteristic and informative words for that particular topic. These words provide insights into the primary themes and content represented by the cluster, effectively inferring the topics present in the Quran-UTM dataset.
4.6 Evaluation Measures
This study uses two well-known topic modeling evaluation metrics—
coherence and
diversity—to evaluate the quality and performance of topic models [Lau et al.
2014; Röder et al.
2015; Terragni et al.
2021]. The coherence measure quantifies the interpretability and meaningfulness of topics by considering the co-occurrence of terms within topics. We employ two coherence metrics, NPMI, [Lau et al.
2014] and
Coherence Value (\(C_V\)), [Röder et al.
2015], and a diversity metric,
Inverted Rank-Biased Overlap (IRBO) [Terragni et al.
2021], to assess the model’s performance. The details of each metric are as follows:
4.6.1 NPMI Score.
The first coherence metric we use to evaluate the topics is NPMI [Bouma
2009], which has been shown to correlate with human judgment [Lau et al.
2014]. It ranges from (–1,1), where –1 signifies that the words never appear together, 0 suggests that their occurrences are entirely independent, and 1 indicates a perfect and consistent co-occurrence of the words. The NPMI score between two words
\(\left(x_{i}, x_{j}\right)\) is calculated as follows:
The probabilities \(p\left(x_{i}\right)\) and \(p\left(x_{i}, x_{j}\right)\) represent the occurrence frequencies of word \(x_i\) and the co-occurrence of word pair \(\left(x_{i}, x_{j}\right)\), respectively. These probabilities are estimated based on their frequencies in a corpus. Here, \(\epsilon\) is added to avoid division by zero. The NPMI score calculates the normalized mutual information between two words and then averages the values across all word pairs in all topics.
4.6.2 \(C_V\) Score.
\(C_V\) is a coherence evaluation metric that assesses the quality of topics generated by topic modeling algorithms. It utilizes a sliding window, a one-set segmentation of the top words, and an indirect confirmation measure involving NPMI and cosine similarity. Given the top
T words of a topic,
\(\left(x_1, x_2, \ldots , x_T\right)\). The calculation of the
\(C_V\) coherence score is as follows:
The NPMI score is computed using Equation (
3). Likewise, the average
\(C_V\) score is obtained by considering the scores of all topics. This approach has been highly influential in evaluating topic quality, as evidenced by extensive comparisons with other commonly used coherence measures. The
\(C_V\) metric yields the best scores that closely align with human evaluations of topic quality (see Röder et al. [
2015] for experimental results).
4.6.3 IRBO Score.
The diversity metric indicates the semantic diversity of the generated topics. Previous studies have proposed different diversity metrics to determine the distinctiveness of the generated topics [Nan et al.
2019; Burkhardt and Kramer
2019; Dieng et al.
2020]. In this study, we use the recently proposed IRBO metric to quantify the diversity of topics [Terragni et al.
2021]. It assigns a score of 0 when the topics are identical and 1 when they are entirely dissimilar. Suppose we have a set
\(\aleph\) containing
T topics. Each topic is represented as a list of words, and the order of words within each list reflects their importance or ranking within the topic. The IRBO of these topics is calculated as follows:
where
\(n= (_2^T)\) represents the number of pairs of lists to be compared. The term
\(RBO\left(l_i, l_j\right)\) refers to the traditional Rank-Biased Overlap measure between two ranked lists
\(l_i\) and
\(l_j\) [Webber et al.
2010]. IRBO enables the comparison of lists even when they do not necessarily share the same items and may not cover all items in a given domain. When two lists (topics) share common words, the IRBO score is lower if the shared words are found at the top than when they are located lower in the ranked lists.
6 Discussion and Analysis
In our study, we performed topic modeling on all of these translations to comprehensively understand the variations in their content. For example, we focused on exploring the topic of “Hellfire” across these translations and made an interesting observation. In Qadri’s translation, this topic was found in 126 different places within the Quranic text, whereas in Kanzuliman’s translation, it appeared only 84 times, as shown in Figure
10. It’s important to note that this difference in the frequency of topic occurrence does not imply that these translations are inherently different, as the Quran is considered the word of Allah. However, it does suggest that specific translations may be more explanatory than others.
Upon manual inspection of Qadri’s translation, we found that it provides additional explanations for the translated verses by including parenthetical explanations. This explanatory translation style may contribute to the higher frequency of topics identified in Qadri’s translation. Figure
11 showcases this feature in Qadri’s translation, highlighting how the inclusion of explanations enhances the richness of the translated text. This simple example shows that topic modeling can be effectively utilized for comparative analysis of the Quran and other religious texts.
Tables
4 and
5 present topics from the Kanzuliman and Ahmed Ali translations, respectively. The variations in word choice and phrasing between the two translations reflect subtle differences in interpretation. For instance, Topic 4 in the Kanzuliman translation frequently references specific figures like
and
, highlighting historical or narrative contexts. In contrast,
in Topic 3 of the Ahmed Ali translation is discussed within a broader context, including terms like
(prophets) or
(messengers). Despite addressing divine entities using different terms, such as
,
, and
, the topics maintain similar meanings in context. It can also be observed that Topic 5 in both translations pertains to the same overarching theme, yet the words and their relevance scores differ. It demonstrates that our BERT-based context-aware topic modeling technique effectively captures the nuances of different translations by considering the surrounding context of words. This approach ensures a more accurate and meaningful extraction of themes, preserving the complexity of the text. This method aids scholars in delving deeper into these topics, examining the variations and nuances in translations, and uncovering further insights. Topic modeling facilitates rigorous analysis, allowing researchers to explore religious texts’ diverse interpretations and perspectives. It promotes meaningful dialogue and advances our understanding of these sacred texts.
7 Conclusion and Future Work
Topic modeling in Urdu translations of the Quran presents significant challenges. Manually identifying topics and conducting comprehensive studies across different translations is highly time-consuming due to the overlapping meanings and ideas between verses and chapters. An automatic method for identifying topics from various translations is needed. While topic modeling has been applied to Quranic translations in English, Arabic, and Indonesian languages, similar techniques for Urdu translations of religious texts are not yet available. In this study, we introduce an approach for topic modeling in Urdu translations of the Quran utilizing BERT-based techniques. Our results reveal the potential of BERT embeddings for topic extraction and analysis. We applied the BERTopic model to eight different Urdu translations of the Quran and compared its performance with two traditional models, LDA and NMF, using coherence and diversity metrics. Our findings show that the BERT-based method surpasses these conventional techniques, with an average coherence improvement of 0.03 and a diversity score of 0.83. These results highlight the effectiveness of BERTopic in identifying meaningful topics from Urdu translations of the Holy Quran and contribute to the computational analysis of religious texts, aiding scholarly research in comparative studies of Quranic translations in Urdu. Despite these promising results, there is room for improvement. One area of improvement is the need for a dedicated Urdu-specific BERT model. Monolingual BERT models exist for various languages, but none specifically for the Urdu language. In future work, developing an Urdu-specific BERT model will significantly enhance the results and improve topic modeling performance in Urdu translations.
Amna Zafar, 
” has the highest representation, while “
” is the least represented word.
