betaFeedback

WorksComputer Science

Non-Deterministic Finite Automata for Telugu Letters and Visualizations for Metrical Poetry

Abstract

Syllabic weight is central to Telugu metrical poetry. Syllabically, every letter (aksharam token) is either light (laghuvu) or heavy (guruvu). Poets classify it using a set of pre-defined rules. This paper formalizes aksharam token, the atomic unit of syllabic classification, as a non-deterministic finite automaton (NFA) with the regular expression ⁣:\colon (Ah+Ad+Ap)(ε+Vs)+Am+H+G(ε+H+Vs+Vv)(A_h+A_d+A_p)(ε+V_s)+A_m+H^+G(ε+H+V_s+V_v), where εε denotes a null string, AA denotes a vowel (achchu), comprising hrasvam (AhA_h), deergham (AdA_d), plutham (ApA_p), and miscellaneous (AmA_m), HH a pure consonant (pollu_hallu), GG a vowel diacritic (gunintha_chihnam), VsV_s the sunna (anusvara), and VvV_v the visarga, the latter two being nasal and aspirated sounds respectively. We then present automata diagrams and regular expressions for (i) a necessary condition satisfied by every light letter, and (ii) a sufficient condition for a letter to be heavy. Finally, we propose graphical visualizations for the 9 Telugu poem types.

Introduction

Telugu is a language with a literary history spanning centuries and a rich inventory of letters. We consider the character set of Telugu Vari Sampoorna Pedda Bala Siksha [12], with the modifications shown in Figure 1, which comprises 16 vowels (achchu- hrasvam, deergham, plutham, miscellaneous), 35 pure consonants (pollu_hallu), the pollu, the sunna, and the visarga. The term ’Aksharam Token’ was introduced in [15]: it denotes a single independent unit that can be classified as light (laghuvu) or heavy (guruvu) in Telugu metrical poetry (chandassu). A character sequence is an AksharamToken if and only if it satisfies one of the following conditions:

Telugu character set

Figure 1. Telugu character set: short vowels (hrasvam), long vowels (deergham), prolated vowels (plutham), miscellaneous, pure consonants (pollu_hallu), short and long vowel diacritics (hrasva gunintha_chihnam, deergha gunintha_chihnam), pollu, sunna, and visarga.

  1. A vowel (achchu), whether short (hrasvam), long (deergham), prolated (plutham) or miscellaneous (described in Figure 1).

  2. A vowel combined with a sunna.

  3. A pure consonant (pollu_hallu) combined with a vowel diacritic (gunintha_chihnam), forming a consonant. A consonant cannot carry more than one vowel diacritic at a time.

  4. Two or more pure consonants combined with a vowel diacritic, forming an orthographic cluster. A cluster cannot carry more than one vowel diacritic at a time.

  5. A consonant followed by a pure consonant (pollu_hallu).

  6. A consonant combined with a sunna.

  7. A consonant combined with a visarga.

  8. An orthographic cluster followed by a pure consonant (pollu_hallu).

  1. An orthographic cluster combined with a sunna.

  1. An orthographic cluster combined with a visarga.

Related Work

The rules of Telugu prosody are found in classical treatises such as Chhan’doo vyaakarand-amu by Anjaneyamurthy [2] and Telugu Vari Sampoorna Pedda Bala Siksha [12] by Gajula Satyanarayana. Boddu Sri Pavan et al. [15], released 4,651 poems (padyams) annotated with laghuvu-guruvu sequences, together with a heuristic-based aksharam tokenizer, syllabic weight generator, and a metric ’Chandassu Score’ on which they report 91.73%. Earlier tools include a chandassu-determination and word-prediction application by Raghu Korrapati et al. [5], a Chandassu Recognizer by I. Reddy Sekhar Reddy et al. [10], and VerseMetrics, a Python based analyzer by S Vijay Kumar et al. [14]. Chandojñanam [13] by Terdalkar, Hrishikesh et al. identifies metres from text or images and offers approximate and close matches for erroneous verses. For Urdu, UPMVM [16] by Asia Zaman et al. develops a DFA for each identified Arud meter, matches verses against 290 meters and sub-meters, and reports an average accuracy of 94%. These automata operate at the metre level.

Finite-state methods on Indic scripts have targeted transliteration and morphology. Tina Bögel et al. [3] use a cascade of finite-state transducers to map Urdu and Hindi scripts into a common ASCII transcription for a shared morphological analyzer. Malik et al. [4] apply finite-state transducers to Hindi-Urdu machine transliteration. Hellsten et al. [11] use weighted finite-state transducers for transliterated keyboard input in South Asian languages, deployed for 22 languages in Gboard. Telugu morphological analysis has followed paradigm-based by G. Uma Maheshwar Rao et al. [7] and machine learning based approaches by Sai Kiranmai et al. [6], both at the word level.

A non-deterministic finite automata (NFA), introduced by M. O. Rabin and D. Scott in 1959[1], is a 5-tuple (Q,Σ,δ,q0,F)(Q, \Sigma, \delta, q_0, F). Here QQ is a finite, non-empty set of states, Σ\Sigma is a finite alphabet, δ:Q×(Σ∪{ϵ})→P(Q)\delta: Q \times(\Sigma\cup\{\epsilon\}) \to\mathcal{P}(Q) is the transition function, q0∈Qq_0 \in Q is the start state, and F⊆QF \subseteq Q is the set of accepting states (final states).

Rodolfo Delmonte’s SPARSAR visualization [9] converts the system’s analysis of a poem into ten graphical maps, grouped into phonetic, poetic (rhyme and metrical structure), and semantic views, and uses colour to highlight contrasts between sound classes and word polarities. A. Abdul-Rahman et al. [8] present a user-centered design study that uses rule-based visual mappings to manage a large set of poetic variables while reducing the effort of operating the mapping control panel, and they compare design options for visualizing poems latitudinally. To our knowledge, neither work targets Telugu chandassu, so neither encodes laghuvu and guruvu weights, ganam types, or yati and prasa positions, which our padyam plots show in a single view.

Prior Telugu tools compute syllabic weight in code, prior automata for prosody operate after weights have been assigned. This paper makes four contributions:

  1. High-level NFA: A non-deterministic finite automaton for the aksharam token, with its regular expression.

  2. Prosody-aware NFA: A refined automaton that separates short and long vowels and diacritics, with its regular expression and a regular expression for the orthographic cluster.

  3. Weight conditions: Two conditions derived from the prosody-aware NFA: a necessary condition satisfied by every light token (laghuvu) and a sufficient condition for a heavy token (guruvu).

  4. Padyam plots: A graphical visualization of Telugu poem types that encodes syllabic weight, yati, prasa, and ganam in a single view, making their metrical structure easier to read.

Proposed Methodology for NFA

Aggregating the rules in the Introduction 1, we propose high-level non-deterministic finite automata (NFA) for the aksharam token, shown in Figure 2. Let AA denotes achhu or vowels, comprising hrasvam (AhA_h), deergham (AdA_d), plutham (ApA_p), and miscellaneous (AmA_m). HH the pollu_hallu, GG the gunintha_chihnam, comprising hrasva_gunintha_chihnam (GsG_s), and deergha_gunintha_chihnam (GdG_d), VsV_s the sunna, VvV_v the visarga. Each is a non-empty set of input symbols. The 5-tuple of the NFA is:

NFA for Aksharam Token

Figure 2. NFA for Aksharam Token

  1. Q={S,H1,AT1,AT2,AT3}Q=\{S, H1, AT1, AT2, AT3\}

  2. q0={S}q_0=\{S\}, marked by ’→\to’

  3. Σ=A∪Am∪H∪G∪Vs∪Vv\Sigma=A\cup A_m\cup H\cup G\cup V_s\cup V_v (Here, AmA_m is included separately as it cannot be combined with sunna-VsV_s)

  4. δ\delta: Table 1

  5. F={AT1,AT2,AT3}F=\{AT1, AT2, AT3\}, marked by ’*’

AAAmA_mHHGGVsV_sVvV_v
→S\to SAT1*AT2*H1ϕ\phiϕ\phiϕ\phi
H1ϕ\phiϕ\phiH1AT3*ϕ\phiϕ\phi
AT1ϕ\phiϕ\phiϕ\phiϕ\phiAT2*ϕ\phi
AT3ϕ\phiϕ\phiAT2*ϕ\phiAT2*AT2*

Table 1. NFA: Transition function table for Aksharam Token (where ϕ\phi is empty set)

The regular expression of the NFA is: (Ah+Ad+Ap)(ϵ+Vs)+Am+H+G(ϵ+H+Vs+Vv)(A_h + A_d + A_p)(\epsilon+ V_s) + A_m + H + G(\epsilon+ H + V_s + V_v)

This NFA accepts aksharam tokens, the atomic units of metrical poetry, each of which is classified as light (laghuvu) or heavy (guruvu). Because vowels include both short and long forms, the transition from ‘S’ to ‘AT1’ can be split further. Likewise, because vowel diacritics have short and long forms, the transition from ‘H1’ to ‘AT3’ can be split. The resulting syllabic-weight-aware automaton is shown in Figure 3. Its 5-tuple is:

Prosody-aware NFA for Aksharam Token

Figure 3. Prosody-aware NFA for Aksharam Token

  1. Q={S,H1,H2,AT1,AT2,AT3,AT4,AT5,AT6,AT7,AT8,AT9,AT10,AT11,AT12,AT13,AT14,AT15,AT16,AT17,AT18,AT19,AT20,AT21}Q=\{S, H1, H2, AT1, AT2, AT3, AT4, AT5, AT6, AT7, AT8, AT9, AT10, AT11, AT12, AT13, AT14, AT15, AT16, AT17, AT18, AT19, AT20, AT21\}

  2. q0={S}q_0=\{S\}; marked by ’→\to’

  1. Σ=Ah∪Ad∪Ap∪Am∪H∪Gh∪Gd∪Vs∪Vv\Sigma= A_h \cup A_d \cup A_p \cup A_m \cup H \cup G_h \cup G_d \cup V_s \cup V_v

  1. δ\delta: Table 2

AhA_hAdA_dApA_pAmA_mHHGhG_hGdG_dVsV_sVvV_v
→S\to SAT1*AT4*AT4*AT3*H1ϕ\phiϕ\phiϕ\phiϕ\phi
AT1ϕ\phiϕ\phiϕ\phiϕ\phiϕ\phiϕ\phiϕ\phiAT2*ϕ\phi
AT4ϕ\phiϕ\phiϕ\phiϕ\phiϕ\phiϕ\phiϕ\phiAT5*ϕ\phi
H1ϕ\phiϕ\phiϕ\phiϕ\phiH2AT6*AT10*ϕ\phiϕ\phi
AT6ϕ\phiϕ\phiϕ\phiϕ\phiAT7*ϕ\phiϕ\phiAT8*AT9*
AT10ϕ\phiϕ\phiϕ\phiϕ\phiAT11*ϕ\phiϕ\phiAT12*AT13*
H2ϕ\phiϕ\phiϕ\phiϕ\phiH2AT14*AT18*ϕ\phiϕ\phi
AT14ϕ\phiϕ\phiϕ\phiϕ\phiAT15*ϕ\phiϕ\phiAT16*AT17*
AT18ϕ\phiϕ\phiϕ\phiϕ\phiAT19*ϕ\phiϕ\phiAT20*AT21*

Table 2. Prosody-aware NFA: Transition table for Aksharam Token (where ϕ\phi is empty set)

  1. F={AT1,AT2,AT3,AT4,AT5,AT6,AT7,AT8,AT9,AT10,AT11,AT12,AT13,AT14,AT15,AT16,AT17,AT18,AT19,AT20,AT21}F = \{\mathrm{AT1}, \mathrm{AT2}, \mathrm{AT3}, \mathrm{AT4}, \mathrm{AT5}, \mathrm{AT6}, \mathrm{AT7}, \mathrm{AT8}, \mathrm{AT9}, \mathrm{AT10}, \mathrm{AT11}, \mathrm{AT12}, \mathrm{AT13}, \mathrm{AT14}, \mathrm{AT15}, \mathrm{AT16}, \mathrm{AT17}, \mathrm{AT18}, \mathrm{AT19}, \mathrm{AT20}, \mathrm{AT21}\}; marked by ‘*’

States AT1, AT2, AT3, AT4, AT5 accept a bare vowel or a vowel followed by a sunna. The remaining final states accept tokens that contain pure consonants: AT6 to AT13 accept tokens without, and AT14 to AT21 accept tokens with orthographic cluster. The regular expression of this syllabic weight-aware NFA is: (Ah+Ad+Ap)(ϵ+Vs)+Am+H+(Gh+Gd)(ϵ+H+Vs+Vv)(A_h + A_d + A_p)(\epsilon+ V_s) + A_m + H^+(G_h + G_d)(\epsilon+ H + V_s + V_v). The regular expression of an orthographic cluster is: HH+(Gh+Gd)(ϵ+H+Vs+Vv)H H^+(G_h + G_d)(\epsilon+ H + V_s + V_v)

From the syllabic weight-aware NFA in Figure 3, we derive two propositions:

Proposition 1 (necessary condition for laghuvu): Every light syllabic weight (laghuvu-I) aksharam token has the form Ah+H+GhA_h + H^{+}G_h.

Proposition 2 (sufficient condition for guruvu): Every aksharam token of the form AhVs+(Ad+Ap)(ϵ+Vs)+Am+H+((Gh+Gd)(H+Vs+Vv)+Gd)A_hV_s + (A_d + A_p)(\epsilon+ V_s) + A_m + H^{+}((G_h + G_d)(H + V_s + V_v) + G_d) is heavy (guruvu-U).

As described in the classical texts [2], [12], the weight of an aksharam token also depends on the token that follows it. If the following token is an orthographic cluster, the current token is heavy, regardless of its intrinsic weight. Proposition 1 is derived from the final states AT1, AT6, and AT14 (light grey in Figure 3). Proposition 2 is derived from all other final states (light green).

PadyamPlot Visualization

We propose a visualization plot for Telugu metrical poetry (padyam), in which each cell represents one aksharam token and its color conveys syllabic weight, position, and ganam type (sequence of syllables). The paadams (lines) of the poem run from top to bottom. Within each line, the ganams (syllabic sequences) run from left to right. Each ganam (syllabic sequence) is drawn as a vertical block whose cells hold the syllabic weights of its aksharam tokens, read from bottom to top.

Weight and position (cell fill and hatching):

  1. An empty cell denotes a light token (laghuvu, I).

  2. A shaded cell (light green) denotes a heavy token (guruvu, U).

  3. Vertical hatching marks a prasa position. It is applied to an empty cell for a light token and to a shaded cell for a heavy token.

  4. Horizontal hatching marks a yati position, applied in the same way.

  5. Inclined hatching marks a position that is prasa yati, applied in the same way.

Ganam type (cell border):

  1. A thick border denotes a Surya ganam.

  2. A dashed border denotes an Indra ganam.

In Figure 4, we present these plots for the following 9 poem types: Vutpalamaala, Champakamaala, Saardulamu, Mattebhamu, Mattakokila, Kandamu, Teytageethi, Aataveladi, and Seesamu, with the dataset mentioned in the Appendix 7.

Vutpalamaala visualization

(a) Vutpalamaala

Champakamala visualization

(b) Champakamala

Saardulamu visualization

(c) Saardulamu

Mattebhamu visualization

(d) Mattebhamu

Mattakokila visualization

(e) Mattakokila

Kandamu visualization

(f) Kandamu

Teytageethi visualization

(g) Teytageethi

Aataveladi visualization

(h) Aataveladi

Seesamu visualization

(i) Seesamu

Figure 4. PadyamPlot Visualization

Conclusion

We presented non-deterministic finite automata (NFAs) and regular expressions for the aksharam token at two levels: high-level and syllabic-weight-aware. We formulated a regular expression for the orthographic cluster. From the weight-aware automaton, we derived two propositions: a necessary condition satisfied by every light token (laghuvu) and a sufficient condition for an aksharam token to be heavy (guruvu). We also proposed padyam plots, graphical

visualizations of Telugu metrical poems that encode syllabic weight, yati, prasa, and ganam, etc. We expect these formal models and visualizations to support further computational work on Telugu chandassu.

Acknowledgements

We thank Kamatham Tharun Kumar for proofreading this paper. We are grateful to our parents and family members for their support. This work is completed by the grace of Sri Ramalinga Chowdeswari Devi.

Appendix

For this work, we considered ganams upto length of four aksharam tokens as shown in Table 3, with two functional classes: INDRA Ganams (BHA, RA, THA, NALA, NAGA, SALA) and SURYA Ganams (HA, NA).

Aksharam Token CountGanam NameLaghuvuGuruvu Pattern
1LA|
1GAU
2LAA||
2VA|U
2HAU|
2GAAUU
3NA|||
3SA||U
3JA|U|
3YA|UU
3BHAU||
3RAU|U
3THAUU|
3MAUUU
4NALA||||
4NAGA|||U
4SALA||U|

Table 3. Ganam classification by aksharam token count and laghuvu-guruvu pattern

Every padyam type below consists of four paadalu (lines).

  1. Vutpalamaala: Each paadam has 20 aksharam tokens and the ganam sequence BHA-RA-NA-BHA-BHA-RA-VA. The yati position is the 10th aksharam token. Prasa applies.

  1. Champakamaala: Each paadam has 21 aksharam tokens and the ganam sequence NA-JA-BHA-JA-JA-JA-RA. The yati position is the 11th aksharam token. Prasa applies.

  1. Saardulamu: Each paadam has 19 aksharam tokens and the ganam sequence MA-SA-JA-SA-THA-THA-GA. The yati position is the 13th aksharam token. Prasa applies.

  1. Mattebhamu Each paadam has 20 aksharam tokens and the ganam sequence SA-BHA-RA-NA-MA-YA-VA. The yati position is the 14th aksharam token. Prasa applies.

  1. Mattakokila Each paadam has 18 aksharam tokens and the ganam sequence RA-SA-JA-JA-BHA-RA. The yati position is the 11th aksharam token. Prasa applies.

  1. Kandamu Each paadam contains the ganams from: GAA, BHA, JA, SA, and NALA. The ganams at odd positions (including the first) must be GAA, BHA, SA, or NALA. The last ganam of the second and fourth paadalu must be GAA or SA, and the 3rd ganam of these paadalu must be JA or NALA. Yati applies only to the second and fourth paadalu, at the first aksharam token of the 4th ganam. Prasa applies.

  1. Teytageethi: Each paadam has the ganam sequence SURYA-INDRA-INDRA-SURYA-SURYA. The yati position is the first aksharam token of the 4th ganam of each paadam. In place of yati, the variant prasa yati applies.

  1. Aataveladi: Each paadam has the ganam sequence SURYA-SURYA-SURYA-INDRA-INDRA. The yati position is the first aksharam token of the 4th ganam of each paadam. In place of yati, the variant prasa yati applies.

  1. Seesamu: Each paadam has the ganam sequence INDRA-INDRA-INDRA-INDRA-INDRA-INDRA-SURYA-SURYA and is split into two sub-paadalu of four ganams each. The first sub-paadam contains four INDRA ganams, and the second contains INDRA-INDRA-SURYA-SURYA. The yati position is the first aksharam token of the 3rd ganam of each sub-paadam. In place of yati, the variant prasa yati applies. Prasa is optional.

References

  1. [1]M. O. Rabin and D. Scott. Finite Automata and Their Decision Problems. IBM Journal of Research and Development, 3(2):114-125, 1959.DOI
  2. [2]Meid’icherla Aan’jineiyamurti. Chhan’doo vyaakarand-amu. Digital Library Of India, 1981.
  3. [3]Tina Bö́gel, Mirjam Butt, Annette Hautli, and Sebastian Sulger. Developing a finite-state morphological analyzer for Urdu and Hindi. In Finite-State Methods and Natural Language Processing, 6th International Workshop, FSMNLP 2007, 2008. https://nbn-resolving.org/urn:nbn:de:kobv:517-opus-27155
  4. [4]M G Abbas Malik, Christian Boitet, and Pushpak Bhattacharyya. Hindi Urdu Machine Transliteration using Finite-State Transducers. In Proceedings of the 22nd International Conference on Computational Linguistics (Coling 2008), pages 537-544, Manchester, UK, 2008. Coling 2008 Organizing Committee.
  5. [5]TVVV Prasad Raghu Korrapati. A natural language application to determine ‘chandassu’ (grammar) & predict the right word in telugu poetry. In Allied Academies International Internet Conference, volume 12, pages 70-76, 2010.
  6. [6]Sai Kiranmai, G., Mallika, K., Anand Kumar, M., Dhanalakshmi, V., Soman, K.P. (2010). Morphological Analyzer for Telugu Using Support Vector Machine. In: Das, V.V., Vijaykumar, R. (eds) Information and Communication Technologies. ICT 2010. Communications in Computer and Information Science, vol 101. Springer, Berlin, Heidelberg. https://doi.org/10.1007/978-3-642-15766-0_68
  7. [7]G. Uma Maheshwar Rao, Amba P. Kulkarni, and Christopher M. A Telugu morphological analyzer. In International Telugu Internet Conference Proceedings, 2011.
  8. [8]Alfie Abdul-Rahman, Julie Lein, Katharine Coles, Eamonn Maguire, Miriah Meyer, Martin Wynne, Chris Johnson, Anne Trefethen, Min Chen. Rule-based Visual Mappings with a Case Study on Poetry Visualization. Computer Graphics Forum (EuroVis), 32(3): doi:10.1111/cgf.12125, 2013.
  9. [9]Rodolfo Delmonte. Visualizing Poetry with SPARSAR – Visual Maps from Poetic Content. In Proceedings of the Fourth Workshop on Computational Linguistics for Literature, pages 68-78, Denver, Colorado, USA. Association for Computational Linguistics, 2015.DOI
  10. [10]M.Humera Khanam I. Reddy Sekhar Reddy. Chandhassu recognizer for telugu poems. International Journal of Computer Science and Information Technologies, 7(3):1425–1434, 2016.
  11. [11]Lars Hellsten, Brian Roark, Prasoon Goyal, Cyril Allauzen, Françoise Beaufays, Tom Ouyang, Michael Riley, and David Rybach. Transliterated Mobile Keyboard Input via Weighted Finite-State Transducers. In Proceedings of the 13th International Conference on Finite State Methods and Natural Language Processing (FSMNLP 2017), pages 10–19, Umeå, Sweden, 2017. Association for Computational Linguistics.
  12. [12]Gajula Satyanarayana. Chandassu. Telugu Vari Sampoorna Pedda Bala Siksha Part 2, pages 243–248, 2019.
  13. [13]Hrishikesh Terdalkar and Arnab Bhattacharya. Chandojñanam: A Sanskrit Meter Identification and Utilization System. In Proceedings of the Computational Sanskrit & Digital Humanities: Selected papers presented at the 18th World Sanskrit Conference, pages 113–127, Canberra, Australia (Online mode), 2023. Association for Computational Linguistics.
  14. [14]S Vijay Kumar Nelli Ramesh. Versemetrics: Telugu chandas analyzer with python. International Journal of Scientific Research in Engineering and Management, 8(11), 2024.
  15. [15]Boddu Sri Pavan and Boddu Swathi Sree. Computational Social Linguistics for Telugu Cultural Preservation: Novel Algorithms for Chandassu Metrical Pattern Recognition. arXiv preprint, 2025. https://arxiv.org/abs/2510.01233
  16. [16]A. Zaman, Z. U. Din, S. Iqbal, and A. A. Shuhail. UPMVM: A Metrics Verification Model for Urdu Poetry. International Journal of Interactive Multimedia and Artificial Intelligence, vol. 9, no. 6, pp. 51-75, 2026, http://doi.org/10.9781/ijimai.2025.09.001

Paper details

vlnc:W_ryf97u8b
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 9

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Computer Science

Not recorded for this paper

View statement
Looking for related papers…
Contents