sangamtamil.com began as an M.Tech dissertation at BITS Pilani: teach a machine to tell which of the five landscapes an akam poem belongs to, using the same cues a scholar uses — the flowers, birds, trades and places named in the verse — and turn what it learns into a knowledge graph anyone can browse. This page summarises the work; the full report is embedded below.
These figures are those of the dissertation as examined in February 2026, and describe the akam corpus the classifier was built on. The site has grown since: the knowledge graph now indexes 1,483 entities across the Sangam anthologies, and the edition covers 47 texts from Tolkappiyam to the epics. Nothing on this page is updated to flatter the report.
An akam poem never names its landscape. A glory lily, a hill and the god Murugan say kurinji; a seagull and salt-pans say neytal. That code lives in Tolkappiyam and in readers' heads, not in any machine-readable form — and general-purpose Tamil language tools, trained on news and social media, fail on two-thousand-year-old vocabulary.
The first classifier was the textbook one: TF-IDF over ten thousand word n-grams and a kernel SVM. It managed a macro F1 of 0.52, and the smallest class, kurinji, collapsed to 0.32. Dropping the "confusing" marutham class did not help — the problem was the features, not a class.
The fix came from literary theory. Tolkappiyam already says what defines a landscape: its karupporul, the flora, fauna, land, occupations and food native to it. Keeping only those entity words in each verse and discarding everything else aligned the feature space with the way a scholar reads. With 145 curated entities the F1 rose to 0.68; with iterative expansion it reached 0.82. As a bonus, every decision became explainable: it can be traced to named cultural entities.
IndicBERT v2 was fine-tuned on the same task in four configurations, each taking about two days on a laptop GPU. The best reached 58.6%. With 1,891 verses there is not enough data to re-teach a twelve-layer encoder an archaic vocabulary that its tokenizer breaks into meaningless fragments, and a transformer's weights offer none of the entity-discovery value the SVM's weights do. The dissertation kept the entity-based SVM and recorded BERT as future work.
Five layers, each improvable on its own. The loop in the middle is the heart of it: the classifier proposes new entities, a human accepts or rejects them, the lexicon grows, the classifier is retrained.
All seven akam texts scraped from Project Madurai — Ainkurunuru, Kuruntokai, Natrinai, Akananuru, Kalittokai, and the two akam idylls Kurinjippattu and Mullaippattu split into verses. Metadata, numbering and commentary stripped; Unicode normalised. No lemmatisation: the inflected poetic forms carry the signal.
500 verses hand-verified, 100 per thinai, for a balanced training and test set. A 176-term seed lexicon of citation-verified flora, fauna, occupations and foods curated from scholarly sources.
Five one-vs-rest linear SVMs, one per thinai, over TF-IDF of entity tokens only. A verse goes to the thinai with the highest decision score; a prediction is trusted for the next step only when score and confidence both clear a threshold.
On verses predicted with confidence above 0.95, terms are scored by SVM coefficient × summed TF-IDF; those above 0.7 and seen at least three times are proposed. A reviewer checks each for cultural relevance, correct thinai and being a real word rather than a suffix or fragment. Sixteen lexicon versions later: 417 entities, 83% acceptance.
A lightweight OWL ontology — Thinai, Verse, Entity, Category — populated from the corpus, every claim traceable to the verse it came from. Exported, then rendered as this website with its knowledge-graph explorer and the Thinai Quest game.
On the balanced 500-verse ground truth. The error profile matches the poetry: the seashore is unmistakable; the wasteland, which Tolkappiyam treats as any landscape withered by drought, borrows everyone else's symbols.
| Thinai | Precision | Recall | F1 |
|---|---|---|---|
| Kurinji குறிஞ்சி | 0.87 | 0.82 | 0.84 |
| Mullai முல்லை | 0.84 | 0.76 | 0.80 |
| Marutham மருதம் | 0.83 | 0.77 | 0.80 |
| Neytal நெய்தல் | 0.88 | 0.87 | 0.87 |
| Palai பாலை | 0.74 | 0.85 | 0.79 |
| Overall | 0.83 | 0.81 | 0.82 |
Neytal is easiest: sea, fishermen, boats and marine life are distinct, consistent cues. Kurinji benefits from its characteristic mountain flora, fauna and terrain. Marutham and mullai share farmland-and-pasture imagery yet stay balanced at 0.80. Palai is hardest — recall is high at 0.85 but precision drops to 0.74, because the wasteland shares its symbols with the landscapes it withers.
Verses per thinai in the full corpus: palai 27.9%, kurinji 27.0%, neytal 18.1%, marutham 13.6%, mullai 13.4%. The natural imbalance was kept, and the two minority classes drove the targeted entity expansions.
Deliberately minimal: four classes, four relationships, and a rule that an entity enters the graph only on double confirmation — a high-confidence thinai classification and a significant SVM weight.
| Graph metric | Value |
|---|---|
| Thinai nodes | 5 |
| Entity nodes | 417 |
| Verse nodes | 1,891 |
| Verse → Thinai edges | 1,891 |
| Entity → Thinai edges | 885 |
| Entity → Verse edges | 19,174 |
| Density | 0.0082 |
| Average degree | 18.98 |
| Maximum degree | 1,117 |
A two-thousand-year-old grammar outperformed ten thousand statistical n-grams. When a corpus is symbolic, encode the symbols.
Removing a class, cleaning too aggressively (versions 5–9 fell to about 78%), and the BERT runs each closed a door and pointed to the right one.
A filtered set of 237 entities matched a 1,000-term set within 1%. The signal lives in a small core of concrete entities.
Sixteen lexicon versions with per-version metadata and a changelog made every gain reproducible and every regression explainable.
The SVM's weights were both the classifier and the entity-discovery engine. A black box would have delivered neither the graph nor the 83% acceptance rate.
Publishing the site and the game turned a dissertation into a resource students, scholars and non-Tamil speakers can actually use.
Twenty-one pages compiled in September 2026 from the dissertation record (final report version 7.0, January 2026). Chapters 1–3 set out the problem, the literature and the design; 4–6 the data, method and results; 7–8 the graph and this platform; 9–10 the journey, limitations and future work.
Download the PDF (11 MB) · Open in a new tab
The report's future-work list asked for broader coverage and multilingual access. That is what this site became after the viva: every Sangam text with its colophons, Tolkappiyam, Thirukkural, the eighteen minor works and the epics, with public-domain English translations that name their translators, and the knowledge graph opened up to every landscape and every verse.