DH 2026

Daejeon, July 27–31

Wed, July 2909:00–10:30S025101-102
Long Paper

Anvay: A Bengali Topic Modelling Dashboard

Vinayak Das Gupta
Shiv Nadar Institution of Eminence, India · vinayak.dasgupta@snu.edu.in

Digital humanities workflows remain anchored in infrastructures designed for English and other Latin-script languages.

Domenico Fiormonte’s early critique of the digital humanities drew attention to the linguistic hierarchies embedded in the technical foundations of the field. He demonstrated how core infrastructural standards encoded a bias toward English and other Latin-script languages. See Domenico Fiormonte, “Towards a Cultural Critique of the Digital Humanities,” Historical Social Research 37, no. 3 (2012): 59–76.

Standard pipelines that are applied to an abugida

An abugida is a type of writing system in which each basic character represents a consonant with an inherent vowel, and additional marks modify that vowel.

script like Bengali fragment, split essential clusters due to zero-width joiners, detach vowel signs, and produce unreadable results due to the proliferation of inflected forms(Sarkar and Bandyopadhyay 2005; Rust et al. 2021). What appears as a “topic” is often a computational residue of orthographic mishandling.

This paper introduces Anvay, the first open-source topic-modelling environment built specifically for Bengali and grounded in the linguistic structure of the script.

To my knowledge, there is currently no topic‑modelling environment designed for Bengali that integrates preprocessing, modelling, and visualization in a reproducible pipeline. Anvay v2.0 also includes experimental support for Hindi, Tamil, and English, allowing cross‑script comparison and adaptation to other abugidas.

Rather than retrofitting English workflows, Anvay begins with Unicode-stable normalisation, conjunct preservation, joiner management, morphology-aware filtering, and extensive stopword removal suited to Bengali syntax. This produces legible topics that do not emerge in English-first pipelines. But Anvay is more than a modelling tool: it is a pedagogical space in which the assumptions of computational methods become visible. Every stage of preprocessing and modelling can be adjusted and rerun.

Anvay is available at: https://github.com/vinayakdasgupta/anvay

Topic modelling has always been a heuristic rather than a descriptive device. In the humanities, scholars have long argued that models reveal patterns only insofar as they are shaped by corpus construction and methodological choice. Jockers (2013) demonstrated how modelling can extend literary interpretation, while Underwood (2019) showed how preprocessing decisions can radically alter a model’s shape. Drucker (2014) reminded us that computational representations encode epistemological assumptions. A topic model becomes a reading practice encoded in software.

DH has responded by developing a rich ecosystem of tools. Command-line systems such as MALLET (McCallum 2002) and Gensim (Řehůřek and Sokja 2010) allow expert control over modelling pipelines; interface-based tools like Voyant (Rockwell and Sinclair n.d.), dfr-browser (Goldstone 2016) and jsLDA (Mimno n.d.) make topic exploration accessible to those uncomfortable with coding. Visualisation frameworks such as LDAvis (Sievert and Shirley 2014) and Topicwizard (Kardos et. al. 2025) allow users to read the output, while more recent embedding-based approaches such as BERTopic (Grootendorst 2022) promise semantic refinement.

While transformer-based models have shown to be more accurate, the reasons not to choose such an approach are twofold: the first is to create a verifiable system that moves away from black-box operations. Second, this is designed as a lightweight, browser-based teaching tool for classrooms that can run on modest hardware.

However, their English-first design is not necessarily suited for handling the Bengali script. Anvay provides an environment that takes the user from preprocessing decisions to visualizations in a single application.

Anvay is implemented as a lightweight Flask-based environment for Bengali topic modelling, built around script-sensitive preprocessing and interpretive visualisation. The system incorporates a custom Unicode normalisation pipeline (resolving conjunct consonants, zero-width joiners, and diacritic inconsistencies), a Bengali-specific tokenizer that preserves orthographic clusters, and a dictionary-driven stemmer tailored to Bengali verbal and nominal morphology. Topic modelling is performed via a modified version of gensim’s LDA (Blei et. al. 2003) with full parameter exposure, while interpretability is supported through multiple visualization libraries and document-level drilldowns linking topics back to sentences. Anvay’s workflow—normalisation, tokenisation, filtering, modelling, and visualisation—is designed for a browser interface (Figure 1). It performs all Bengali-specific processing automatically while allowing users to adjust parameters that shape the model. Because it avoids transformer-scale dependencies, it can run on modest machines and be deployed in classrooms without specialised hardware.

Figure 1: Workflow for Anvay

The pedagogical value of such visibility is considerable. As students interactively adjust tokenisation rules or stemming parameters, they see the effects across the resulting topics. This real-time, causal feedback demystifies the modelling pipeline and fosters a deeper understanding of how meaning is computationally constructed. In Indian classrooms—where DH training often requires overcoming both linguistic alienation from English-centric tools and technical hesitation—the ability to work directly and computationally in Bengali expands the possibilities of multilingual DH. This transparency is central to Anvay's design: as topic modelling becomes a visible and configurable process rather than a black box, it is supported by extensive documentation that includes interactive learning activities, parameter guides, and a step-by-step walkthrough. Users can experiment with a simulated topic model in both English and Bengali to grasp core concepts like stopwords and word-topic assignment. A detailed walkthrough with a sample corpus demonstrates how filtering choices shape thematic visibility, reinforcing the tool’s role as a critical learning environment. This documentation actively guides users from linguistic intuition to computational interpretation, grounding DH pedagogy in hands-on inquiry.

To illustrate these claims, the presentation will demonstrate the various aspects of the software. It will show the modelling in action using Rabindranath Tagore’s literary works. These texts, dense with compounds and inflectional variation, quickly reveal the limits of standard pipelines (see Table 1). For instance, the results from the analysis of 65 of Tagore’s dramatic texts provides interesting reading: a stock Gensim model produces a topic dominated by malformed fragments (চল, মহ, মর, পথ). These word fragments are split due to mishandling of zero-width joiners. Anvay produces a semantically coherent cluster that surfaces recurring oppositions in Tagore’s dramaturgy. The words that form the topic show interesting relationships: we can see contrasts of evil versus devotion (রাক্ষস vs. নিতাই), punishment versus liberation (দণ্ডে vs. লঘু), and ritual object versus sacrificial act (প্রতিমা vs. পিসে). The co-occurrence of বিপ্র with a named female figure (মালিনীর) hints at the entanglements of caste, gender, and ritual authority. These are recurring themes in Tagore’s dramatic works, where devotion collides with social hierarchy, and where ritual is both a burden and a release.  The coherence of the resulting topics will be assessed through the human interpretability of top-term lists and their contextual fit within the plays, moving beyond purely algorithmic coherence or perplexity scores. This human-centric evaluation aligns with the critical DH practice of treating models as heuristic devices.

Stock Gensim outputAnvay output
চল, মহ, মর, পথ, দয়, মত, ধন, সম, দন, রহ রাক্ষস (demon), দণ্ডে (punishment), প্রতিমা (idol), গহনা (ornament), বিপ্র (Brahmin or sage), মালিনীর (of Malini), নিতাই (devotional figure), লঘু (liberation or trivial), পিসে (crushed), লাগিবে (will be needed)

Table 1: Stock Tokenizer vs Custom Tokenizer

In conclusion, Anvay demonstrates that true translingual engagement in DH requires moving beyond surface translation to script-aware infrastructural reinvention. By making its pipeline transparent and interactive, it fosters a critical, hands-on pedagogy of modelling. This paper will present, through examples and demonstration, the theoretical rationale, architectural design, and pedagogical outcomes of Anvay, arguing for a script-centric paradigm in the development of multilingual digital humanities tools.

References
  1. Blei, D. M., Ng, A. Y., & Jordan, M. I. 2003. ‘Latent Dirichlet allocation.’ Journal of Machine Learning Research, no. 3, 993–1022.
  2. Drucker, Johanna. 2014. Graphesis: Visual Forms of Knowledge Production. Cambridge, MA: Harvard University Press.
  3. Fiormonte, Domenico. 2012. "Towards a Cultural Critique of the Digital Humanities." Historical Social Research 37, no. 3: 59–76.
  4. Goldstone, Andrew. 2016. Dfr-Browser. https://agoldst.github.io/dfr-browser/.
  5. Grootendorst, Maarten. 2022. “BERTopic: Neural Topic Modeling with a Class-Based TF–IDF Procedure.” arXiv preprint arXiv:2203.05794.
  6. Jockers, Matthew L. 2013. Macroanalysis: Digital Methods and Literary History. Urbana: University of Illinois Press.
  7. Kardos, Márton, Kenneth Enevoldsen, and Kristoffer Nielbo. 2025. “topicwizard – a Modern, Model-Agnostic Framework for Topic Model Visualization and Interpretation.” In Proceedings of the 8th International Conference on Natural Language and Speech Processing (ICNLSP-2025), edited by Mourad Abbas, Tariq Yousef, and Lukas Galke, 19–27. Odense, Denmark: Association for Computational Linguistics. https://aclanthology.org/2025.icnlsp-1.3/.
  8. McCallum, Andrew Kachites. 2002. MALLET: A Machine Learning for Language Toolkit. http://mallet.cs.umass.edu.
  9. Mimno, David. n.d. jsLDA: In-Browser Topic Modelling. https://mimno.infosci.cornell.edu/jsLDA/.
  10. Řehůřek, Radim, and Petr Sojka. 2010. “Software Framework for Topic Modelling with Large Corpora.” In Proceedings of the LREC 2010 Workshop on New Challenges for NLP Frameworks, 45–50. Valletta, Malta: ELRA. http://is.muni.cz/publication/884893/en.
  11. Rockwell, Geoffrey, and Stéfan Sinclair. n.d. Voyant Tools. https://voyant-tools.org/.
  12. Rust, Phillip, Jonas Pfeiffer, Ivan Vulić, Sebastian Ruder, and Iryna Gurevych. 2021. “How Good Is Your Tokenizer? On the Monolingual Performance of Multilingual Language Models.” In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), 3118–3135. Online: Association for Computational Linguistics.
  13. Sarkar, Sudeshna, and Sivaji Bandyopadhyay. 2005. “Morphology of Bangla Verbs: A Computational Perspective.” Proceedings of the ACL Workshop on South Asian Languages.
  14. Sievert, Carson, and Kenneth R. Shirley. 2014. “LDAvis: A Method for Visualizing and Interpreting Topics.” In Proceedings of the Workshop on Interactive Language Learning, Visualization, and Interfaces, 63–70. Association for Computational Linguistics. http://www.aclweb.org/anthology/W14-3110.
  15. Underwood, Ted. 2019. Distant Horizons: Digital Evidence and Literary Change. Chicago: University of Chicago Press.