Eigenmood Space: Uncertainty-Aware Spectral Graph Analysis of Psychological Patterns in Classical Persian Poetry

Kourosh Shahnazari; Seyed Moein Ayyoubzadeh; Mohammadali Keshtparvar

cs.CLFeb 18, 2026

Eigenmood Space: Uncertainty-Aware Spectral Graph Analysis of Psychological Patterns in Classical Persian Poetry

Authors: Kourosh Shahnazari, Seyed Moein Ayyoubzadeh, Mohammadali Keshtparvar

Organizations: Sharif University of Technology, Tehran, Iran. · Amirkabir University of Technology, Tehran, Iran.

Abstract

Classical Persian poetry is a historically sustained archive in which affective life is expressed through metaphor, intertextual convention, and rhetorical indirection. These properties make close reading indispensable while limiting reproducible comparison at scale. We present an uncertainty-aware computational framework for poet-level psychological analysis based on large-scale automatic multi-label annotation. Each verse is associated with a set of psychological concepts, per-label confidence scores, and an abstention flag that signals insufficient evidence. We aggregate confidence-weighted evidence into a Poet $\times$ Concept matrix, interpret each poet as a probability distribution over concepts, and quantify poetic individuality as divergence from a corpus baseline using Jensen--Shannon divergence and Kullback--Leibler divergence. To capture relational structure beyond marginals, we build a confidence-weighted co-occurrence graph over concepts and define an Eigenmood embedding through Laplacian spectral decomposition. On a corpus of 61{,}573 verses across 10 poets, 22.2% of verses are abstained, underscoring the analytical importance of uncertainty. We further report sensitivity analysis under confidence thresholding, selection-bias diagnostics that treat abstention as a category, and a distant-to-close workflow that retrieves verse-level exemplars along Eigenmood axes. The resulting framework supports scalable, auditable digital-humanities analysis while preserving interpretive caution by propagating uncertainty from verse-level evidence to poet-level inference.

Figures & tables

Figure 1: Distribution of confidence scores across all label assignments.

Concept	Weighted mass	Share ( $P_{0}$ )
Melancholia	15462.2	0.307
Emotional Dependency	9962.4	0.198
Romantic Obsession	8564.9	0.170
Self-Destructive Idealization	5900.7	0.117
Spiritual Narcissism	3955.4	0.078
Ambivalent Attachment	2483.0	0.049

Table 1: Global confidence-weighted concept distribution.

Figure 2: Global psychological concept distribution (confidence-weighted).

Concept	Base	$\tau=0.5$	$\tau=0.7$
Melancholia	0.307	0.306	0.302
Emotional Dependency	0.198	0.198	0.206
Romantic Obsession	0.170	0.170	0.202
Self-Destructive Idealization	0.117	0.117	0.106
Spiritual Narcissism	0.078	0.078	0.081
Ambivalent Attachment	0.049	0.049	0.036

Table 2: Global concept shares under confidence thresholding (non-abstained verses only).

Figure 3: Abstention rate by poet.

Figure 4: Poet individuality measured as Jensen–Shannon divergence from baseline.

Figure 5: Association between abstention rate and poet individuality (

D_{JS}

Poet	Verses	Abstain	Avg. conf.	$D_{JS}$	Top concepts
Khaghani	17292	0.357	0.701	0.0179	Melancholia (0.316) Spiritual Narcissism (0.164)
Jahan	12299	0.065	0.710	0.0250	Melancholia (0.319) Emotional Dependency (0.305)
Saadi	6864	0.137	0.708	0.0106	Romantic Obsession (0.239) Melancholia (0.238)
Vahshi	5727	0.289	0.698	0.0085	Melancholia (0.327) Self-Destructive Idealization (0.152)
Parvin	5573	0.384	0.695	0.0459	Melancholia (0.370) Self-Destructive Idealization (0.190)
Hafez	5221	0.182	0.693	0.0035	Melancholia (0.294) Romantic Obsession (0.188)

Table 3: Poet-level summary: verses, abstention, mean confidence, divergence, and dominant concepts (shares from

P_{i}

Figure 6: Poet $\times$ Concept profiles as confidence-weighted distributions $P_{i}(c)$ (darker indicates higher share).

Poet	Largest positive lifts	Largest negative lifts
Khayyam	Melancholia $0.232$ Identity Fragmentation $0.101$ Spiritual Narcissism $0.022$	Emotional Dependency $-0.169$ Romantic Obsession $-0.116$ Ambivalent Attachment $-0.040$
Parvin	Self-Destructive Idealization $0.073$ Melancholia $0.063$ Identity Fragmentation $0.062$	Romantic Obsession $-0.126$ Emotional Dependency $-0.112$ Ambivalent Attachment $-0.006$
Jahan	Emotional Dependency $0.107$ Romantic Obsession $0.027$ Melancholia $0.013$	Spiritual Narcissism $-0.063$ Self-Destructive Idealization $-0.050$ Identity Fragmentation $-0.029$

Table 4: Concept lifts for the three most distinctive poets, defined as

\Delta_{i}(c)=P_{i}(c)-P_{0}(c)

(positive indicates over-representation relative to baseline).

Poet	Baseline	$\tau=0.5$	$\tau=0.7$
Khayyam	0.0901	0.0900	0.1011
Parvin	0.0459	0.0460	0.0564
Jahan	0.0250	0.0251	0.0282
Khaghani	0.0179	0.0180	0.0227
Saadi	0.0106	0.0108	0.0160
Vahshi	0.0085	0.0085	0.0111

Table 5: Jensen–Shannon divergence under confidence thresholding. Baseline uses all non-abstained labels; thresholded variants include only labels with

p_{v,c}\geq\tau

Poet	Non-abstained verses	$D_{JS}$ mean [2.5%, 97.5%]	EM2 mean [2.5%, 97.5%]
Khayyam	219	0.094 [0.072, 0.122]	-0.058 [-0.085, -0.028]
Parvin	3432	0.046 [0.041, 0.051]	-0.027 [-0.034, -0.019]
Jahan	11502	0.025 [0.024, 0.027]	0.007 [0.005, 0.009]
Khaghani	11111	0.018 [0.016, 0.020]	-0.011 [-0.014, -0.008]
Saadi	5927	0.011 [0.010, 0.012]	-0.003 [-0.006, 0.000]
Vahshi	4073	0.009 [0.007, 0.011]	0.007 [0.002, 0.012]

Table 6: Bootstrap uncertainty intervals (200 replicates) for individuality and EM2. Intervals are 2.5–97.5% percentiles.

Figure 7: Top concept co-occurrence network (top 20 edges by confidence-weighted co-activation).

Figure 8: Eigenmood embedding in two dimensions (EM2 vs EM3). Point size scales with verse count and color encodes $D_{JS}$ .

Concept pair	Edge weight
Emotional Dependency – Melancholia	4739.1
Emotional Dependency – Romantic Obsession	2650.9
Melancholia – Romantic Obsession	1981.4
Melancholia – Self-Destructive Idealization	1726.7
Ambivalent Attachment – Melancholia	1124.0
Emotional Dependency – Self-Destructive Idealization	770.6

Table 7: Top confidence-weighted concept co-occurrence edges.

Axis	Largest-magnitude loadings (concept, value)
EM1	Spiritual Narcissism $0.931$ Ambivalent Attachment $-0.217$ Melancholia $-0.135$
EM2	Internal Projection $0.837$ Identity Fragmentation $-0.531$ Self-Destructive Idealization $-0.079$
EM3	Identity Fragmentation $-0.707$ Ambivalent Attachment $0.537$ Internal Projection $-0.381$

Table 8: Largest-magnitude concept loadings for the first three Eigenmood axes (filtered 8-concept graph).

Poet	$D_{JS}$	Abstain rate	Verses
Khayyam	$0.0901$	$0.385$	$356$
Parvin	$0.0459$	$0.384$	$5573$
Jahan	$0.0250$	$0.065$	$12\,299$
Khaghani	$0.0179$	$0.357$	$17\,292$
Saadi	$0.0106$	$0.137$	$6864$
Vahshi	$0.0085$	$0.289$	$5727$

Table 9: Ranking of poets by divergence from the global baseline (

D_{JS}

Figure 9: Cross-poet Eigenmood embedding (EM2 vs EM3). Labels are offset outside markers; point size scales with verse count; color encodes $D_{JS}$ .

Figure 10: EM1 high-side retrieval exemplar for Khaghani.

Figure 11: EM1 low-side retrieval exemplar for Jahan.

Figure 12: EM2 high-side retrieval exemplar for Eraghi.

Figure 13: EM2 low-side retrieval exemplar for Khayyam.

Figure 14: EM3 high-side retrieval exemplar for Jahan.

Figure 15: EM3 low-side retrieval exemplar for Parvin.

Figure 16: Divergence-linked retrieval exemplar for Khayyam.

Figure 17: Divergence-linked retrieval exemplar for Hafez.

Figure 18: Divergence-linked retrieval exemplar for Parvin.

Figure 19: Divergence-linked retrieval exemplar for Shahriar.

Concept	$p_{o}$	$p_{e}$	$\kappa$	Pos A	Pos B
Ambivalent Attachment	0.980	0.931	0.712	16	20
Emotional Dependency	0.978	0.744	0.914	74	77
Idealization	0.990	0.986	0.282	5	2
Identity Fragmentation	0.984	0.909	0.825	24	24
Internal Projection	0.984	0.931	0.770	19	17
Melancholia	0.962	0.660	0.888	109	108

Table 10: Inter-annotator agreement by concept on the 500-verse validation sample. We report observed agreement

p_{o}

, expected agreement

p_{e}

under annotator marginals, and Cohen’s

\kappa

. Pos A and Pos B report the number of verses labeled positive by Annotator A and Annotator B.

Concept	#Pred	#Correct	Precision
Ambivalent Attachment	21	15	0.714
Emotional Dependency	81	70	0.864
Idealization	3	2	0.667
Identity Fragmentation	27	19	0.704
Internal Projection	18	16	0.889
Melancholia	132	105	0.795

Table 11: Model agreement with a two-annotator adjudicated reference on the validation sample. Precision is computed over model-predicted label instances.

Concept	Precision	Recall	$F_{1}$	Support
Ambivalent Attachment	0.714	0.652	0.682	23
Emotional Dependency	0.864	0.864	0.864	81
Idealization	0.667	0.333	0.444	6
Identity Fragmentation	0.704	0.679	0.691	28
Internal Projection	0.889	0.727	0.800	22
Melancholia	0.795	0.890	0.840	118

Table 12: Per-concept precision, recall, and

F_{1}

on the 500-verse validation sample against a two-annotator adjudicated reference. Support reports the number of reference-positive verses for each concept (

|G_{c}|

Bin	Count	Mean conf.	Accuracy	$\|$ gap $\|$
[0.3,0.4)	4	0.327	0.500	0.173
[0.5,0.6)	31	0.509	0.710	0.201
[0.6,0.7)	93	0.673	0.710	0.036
[0.7,0.8)	55	0.751	0.764	0.012
[0.8,0.9)	151	0.843	0.821	0.022
[0.9,1.0)	103	0.944	0.951	0.008

Table 13: Calibration table for model-predicted label instances after temperature scaling. Each bin reports the number of instances, mean scaled confidence, empirical accuracy against the adjudicated reference, and the absolute calibration gap.

Figure 20: Calibration reliability diagram (temperature-scaled confidences) on the validation sample.

$\tau$	Retained	Coverage	Accuracy	Risk
0.3	437	1.000	0.810	0.190
0.5	433	0.991	0.813	0.187
0.7	309	0.707	0.854	0.146
0.8	254	0.581	0.874	0.126
0.9	103	0.236	0.951	0.049

Table 14: Coverage–risk table over temperature-scaled confidence thresholds

\tau

. Coverage is the fraction of retained model-predicted label instances with scaled confidence at least

\tau

, and risk is

1-\mathrm{accuracy}

on the retained set.

Figure 21: Coverage–risk curve on the validation sample as the temperature-scaled confidence threshold

\tau

increases.

Comparison	Spearman $\rho$	$p$ -value	95% bootstrap CI
Base vs abstention-inclusive	0.903	$3.4\times 10^{-4}$	$[0.47,\,1.00]$

Table 15: Rank stability of poet individuality (

D_{JS}

) under abstention modeling. Spearman correlation is computed between poet rankings induced by baseline individuality

D_{JS}(P_{i},P_{0})

and abstention-inclusive individuality

D_{JS}(P_{i}^{\prime},P_{0}^{\prime})

Poet	Abstain rate	$D_{JS}$ (base)	$D_{JS}$ (aug.)	Rank change
Khayyam	0.385	0.0901	0.0850	+0
Parvin	0.384	0.0459	0.0547	+0
Jahan	0.065	0.0250	0.0505	+0
Khaghani	0.357	0.0179	0.0274	+0
Saadi	0.137	0.0106	0.0156	+1
Vahshi	0.289	0.0085	0.0102	+2

Table 16: Selection-bias sensitivity analysis by treating abstention as an explicit category. Rank change is defined as

(\mathrm{rank}_{\mathrm{aug}}-\mathrm{rank}_{\mathrm{base}})

; negative values indicate that a poet becomes more distinctive under abstention-inclusive individuality.

Appendix figures & tables80 assets

Supplementary material from the paper’s appendix.

Appendix

Poet	Verses	Abstain	$D_{JS}$
Khaghani	$17\,292$	$0.357$	$0.0179$

Appendix

Table B1: Core profile metrics for Khaghani .

Poet	EM1	EM2	EM3
Khaghani	$0.092$	$-0.011$	$-0.018$

Appendix

Table B2: Eigenmood coordinates for Khaghani .

Rank	Concept	$P_{i}(c)$
1	Melancholia	$0.316$
2	Spiritual Narcissism	$0.165$
3	Self-Destructive Idealization	$0.151$

Appendix

Table B3: Top concept shares for Khaghani . Concept names are abbreviated in the header; full interpretation appears in the text.

Figure B1: Khaghani: concept distribution

P_{i}(c)

against the global baseline

P_{0}(c)

Figure B2: Khaghani: lift profile

\Delta_{i}(c)=P_{i}(c)-P_{0}(c)

with positive and negative deviations.

Figure B3: Khaghani: bootstrap 95% intervals for the top-3 concepts by

P_{i}(c)

Reference	Score	English translation
EM1 (positive)	$0.838$	Let them call me a second Ptolemy. (Translation ours.)
EM1 (positive)	$0.791$	On my threshold stood hosts of angels in rank. (Translation ours.)
EM1 (positive)	$0.791$	I was a teacher in the heavenly realm; from my obedience came thousands upon thousands of treasures. (Translation ours.)

Appendix

Table B4: Khaghani: top verses on the poet’s most extreme Eigenmood axis (EM1, positive direction).

Reference	Score	English translation
Melancholia	$0.950$	Sorrows drained Khaghani’s blood; return him, at least, pure blood. (Translation ours.)
Melancholia	$0.950$	You made my body and soul a dwelling of grief; in my heart’s garden the lily of sorrow bloomed. (Translation ours.)
Melancholia	$0.900$	My heart has been worn away beneath grief’s tread; my hand is scorched in grief’s fire. (Translation ours.)

Appendix

Table B5: Khaghani: top verses for the leading concept (Melancholia).

Poet	Verses	Abstain	$D_{JS}$
Jahan	$12\,299$	$0.065$	$0.0250$

Appendix

Table B6: Core profile metrics for Jahan .

Poet	EM1	EM2	EM3
Jahan	$-0.068$	$0.007$	$0.031$

Appendix

Table B7: Eigenmood coordinates for Jahan .

Rank	Concept	$P_{i}(c)$
1	Melancholia	$0.319$
2	Emotional Dependency	$0.305$
3	Romantic Obsession	$0.197$

Appendix

Table B8: Top concept shares for Jahan . Concept names are abbreviated in the header; full interpretation appears in the text.

Figure B4: Jahan: concept distribution

P_{i}(c)

against the global baseline

P_{0}(c)

Figure B5: Jahan: lift profile

\Delta_{i}(c)=P_{i}(c)-P_{0}(c)

with positive and negative deviations.

Figure B6: Jahan: bootstrap 95% intervals for the top-3 concepts by

P_{i}(c)

Reference	Score	English translation
EM1 (negative)	$-0.381$	With all your harshness, ill temper, and broken vows, you remain the soul’s companion and my wounded heart’s hope. (Translation ours.)
EM1 (negative)	$-0.333$	He carried off my heart and faith and veiled his face from me; what can I do, faithlessness is the beloved’s habit. (Translation ours.)
EM1 (negative)	$-0.331$	I have neither patience without him nor strength to bear his cruelty; my heart is at its limit, what am I to do? (Translation ours.)

Appendix

Table B9: Jahan: top verses on the poet’s most extreme Eigenmood axis (EM1, negative direction).

Reference	Score	English translation
Melancholia	$0.950$	Not a night passes in the sorrow of your absence that my face is not washed with liver-blood. (Translation ours.)
Melancholia	$0.950$	In your separation my tears run like the Tigris; from grief, my liver’s blood drips from my eyes. (Translation ours.)
Melancholia	$0.900$	I am weary of grief, weary of the world; it holds nothing but sorrow for our sorrow-laden heart. (Translation ours.)

Appendix

Table B10: Jahan: top verses for the leading concept (Melancholia).

Poet	Verses	Abstain	$D_{JS}$
Saadi	$6864$	$0.137$	$0.0106$

Appendix

Table B11: Core profile metrics for Saadi .

Poet	EM1	EM2	EM3
Saadi	$-0.047$	$-0.003$	$0.007$

Appendix

Table B12: Eigenmood coordinates for Saadi .

Rank	Concept	$P_{i}(c)$
1	Romantic Obsession	$0.239$
2	Melancholia	$0.238$
3	Emotional Dependency	$0.224$

Appendix

Table B13: Top concept shares for Saadi . Concept names are abbreviated in the header; full interpretation appears in the text.

Figure B7: Saadi: concept distribution

P_{i}(c)

against the global baseline

P_{0}(c)

Figure B8: Saadi: lift profile

\Delta_{i}(c)=P_{i}(c)-P_{0}(c)

with positive and negative deviations.

Figure B9: Saadi: bootstrap 95% intervals for the top-3 concepts by

P_{i}(c)

Reference	Score	English translation
EM1 (negative)	$-0.329$	The beloved forbids my union; I desist from what he forbids. The one I behold decrees my death; I submit to what he commands. (Translation ours.)
EM1 (negative)	$-0.318$	Rise, fill the cup, and pour, even if poison is in it; from your hand, even poison becomes remedy. (Translation ours.)
EM1 (negative)	$-0.301$	My thought is to pour out my life for you, not to free myself from your lasso. (Translation ours.)

Appendix

Table B14: Saadi: top verses on the poet’s most extreme Eigenmood axis (EM1, negative direction).

Reference	Score	English translation
Romantic Obsession	$0.950$	Wherever I turned my gaze, your image appeared before me on wall and door. (Translation ours.)
Romantic Obsession	$0.900$	We rose, yet your imprint stayed within our breath; wherever you are absent, we cannot remain. (Translation ours.)
Romantic Obsession	$0.900$	Pity the eye that never saw your face, or saw it once and looked again at another. (Translation ours.)

Appendix

Table B15: Saadi: top verses for the leading concept (Romantic Obsession).

Poet	Verses	Abstain	$D_{JS}$
Vahshi	$5727$	$0.289$	$0.0085$

Appendix

Table B16: Core profile metrics for Vahshi .

Poet	EM1	EM2	EM3
Vahshi	$0.020$	$0.007$	$0.021$

Appendix

Table B17: Eigenmood coordinates for Vahshi .

Rank	Concept	$P_{i}(c)$
1	Melancholia	$0.327$
2	Self-Destructive Idealization	$0.152$
3	Emotional Dependency	$0.138$

Appendix

Table B18: Top concept shares for Vahshi . Concept names are abbreviated in the header; full interpretation appears in the text.

Figure B10: Vahshi: concept distribution

P_{i}(c)

against the global baseline

P_{0}(c)

Figure B11: Vahshi: lift profile

\Delta_{i}(c)=P_{i}(c)-P_{0}(c)

with positive and negative deviations.

Figure B12: Vahshi: bootstrap 95% intervals for the top-3 concepts by

P_{i}(c)

Reference	Score	English translation
EM3 (positive)	$0.505$	I remember the day your hem was in my hand: you struck with the dagger, and I kissed your feet. (Translation ours.)
EM3 (positive)	$0.499$	If you know how many birds circle your snare, cast your trap in the plain, but first cast it over our hearts. (Translation ours.)
EM3 (positive)	$0.499$	To strike down thousands unaware, the flower-dagger is hidden beneath the hem. (Translation ours.)

Appendix

Table B19: Vahshi: top verses on the poet’s most extreme Eigenmood axis (EM3, positive direction).

Reference	Score	English translation
Melancholia	$0.900$	Do not ask after my wounded heart; leave me for a moment with my sorrow and affliction. (Translation ours.)
Melancholia	$0.900$	Restless, disordered in mind, constricted in heart, and grief-stricken. (Translation ours.)
Melancholia	$0.850$	I once held a treasure of patience in my heart; the army of sorrow plundered that whole treasury. (Translation ours.)

Appendix

Table B20: Vahshi: top verses for the leading concept (Melancholia).

Poet	Verses	Abstain	$D_{JS}$
Parvin	$5573$	$0.384$	$0.0459$

Appendix

Table B21: Core profile metrics for Parvin .

Poet	EM1	EM2	EM3
Parvin	$0.039$	$-0.027$	$-0.056$

Appendix

Table B22: Eigenmood coordinates for Parvin .

Rank	Concept	$P_{i}(c)$
1	Melancholia	$0.370$
2	Self-Destructive Idealization	$0.190$
3	Spiritual Narcissism	$0.116$

Appendix

Table B23: Top concept shares for Parvin . Concept names are abbreviated in the header; full interpretation appears in the text.

Figure B13: Parvin: concept distribution

P_{i}(c)

against the global baseline

P_{0}(c)

Figure B14: Parvin: lift profile

\Delta_{i}(c)=P_{i}(c)-P_{0}(c)

with positive and negative deviations.

Figure B15: Parvin: bootstrap 95% intervals for the top-3 concepts by

P_{i}(c)

Reference	Score	English translation
EM3 (negative)	$-0.601$	The soul’s radiant palace stands firm; what endurance and permanence can the body’s hut possess? (Translation ours.)
EM3 (negative)	$-0.601$	The head, without the lamp of reason, is trapped in darkness; the body without spirit is scattered like dust. (Translation ours.)
EM3 (negative)	$-0.601$	A manifest tale can never be hidden from the eye of reason. (Translation ours.)

Appendix

Table B24: Parvin: top verses on the poet’s most extreme Eigenmood axis (EM3, negative direction).

Reference	Score	English translation
Melancholia	$0.900$	Sit in a heart that has been cleansed of grief; then gladden me, O comforter of sorrowing hearts. (Translation ours.)
Melancholia	$0.900$	Why must one live so desolate, dead at the morning of life? (Translation ours.)
Melancholia	$0.900$	What shall I sing but grief’s melody? What shall I say to moonlight and dew? (Translation ours.)

Appendix

Table B25: Parvin: top verses for the leading concept (Melancholia).

Poet	Verses	Abstain	$D_{JS}$
Hafez	$5221$	$0.182$	$0.0035$

Appendix

Table B26: Core profile metrics for Hafez .

Poet	EM1	EM2	EM3
Hafez	$0.019$	$0.003$	$-0.027$

Appendix

Table B27: Eigenmood coordinates for Hafez .

Rank	Concept	$P_{i}(c)$
1	Melancholia	$0.294$
2	Romantic Obsession	$0.188$
3	Emotional Dependency	$0.160$

Appendix

Table B28: Top concept shares for Hafez . Concept names are abbreviated in the header; full interpretation appears in the text.

Figure B16: Hafez: concept distribution

P_{i}(c)

against the global baseline

P_{0}(c)

Figure B17: Hafez: lift profile

\Delta_{i}(c)=P_{i}(c)-P_{0}(c)

with positive and negative deviations.

Figure B18: Hafez: bootstrap 95% intervals for the top-3 concepts by

P_{i}(c)

Reference	Score	English translation
EM3 (negative)	$-0.759$	Within my wounded heart I do not know who cries out; I am silent, while he is uproar and lament. (Translation ours.)
EM3 (negative)	$-0.759$	Do not say that, absent heart, I forgot you by mistake. (Translation ours.)
EM3 (negative)	$-0.724$	My heart was a treasury of secrets; fate locked it and handed its key to the stealer of hearts. (Translation ours.)

Appendix

Table B29: Hafez: top verses on the poet’s most extreme Eigenmood axis (EM3, negative direction).

Reference	Score	English translation
Melancholia	$0.900$	My palate turned to poison from grief’s bitterness; remember the cry of those who drink joy. (Translation ours.)
Melancholia	$0.900$	I go to the tavern weeping and seeking redress; perhaps there I may be released from sorrow’s hand. (Translation ours.)
Melancholia	$0.900$	I have endured a love-pain beyond asking; I have tasted the poison of separation beyond asking. (Translation ours.)

Appendix

Table B30: Hafez: top verses for the leading concept (Melancholia).

Poet	Verses	Abstain	$D_{JS}$
Eraghi	$4668$	$0.081$	$0.0040$

Appendix

Table B31: Core profile metrics for Eraghi .

Poet	EM1	EM2	EM3
Eraghi	$0.014$	$0.016$	$-0.010$

Appendix

Table B32: Eigenmood coordinates for Eraghi .

Rank	Concept	$P_{i}(c)$
1	Melancholia	$0.277$
2	Emotional Dependency	$0.241$
3	Romantic Obsession	$0.181$

Appendix

Table B33: Top concept shares for Eraghi . Concept names are abbreviated in the header; full interpretation appears in the text.

Figure B19: Eraghi: concept distribution

P_{i}(c)

against the global baseline

P_{0}(c)

Figure B20: Eraghi: lift profile

\Delta_{i}(c)=P_{i}(c)-P_{0}(c)

with positive and negative deviations.

Figure B21: Eraghi: bootstrap 95% intervals for the top-3 concepts by

P_{i}(c)

Reference	Score	English translation
EM2 (positive)	$0.796$	In vain I sought you through the world; you were already within Iraqi’s soul. (Translation ours.)
EM2 (positive)	$0.754$	If you look to left and right, behind and before, know that before, behind, left, right, and above are all He. (Translation ours.)
EM2 (positive)	$0.712$	He set speech upon our tongues and planted seeking within us. (Translation ours.)

Appendix

Table B34: Eraghi: top verses on the poet’s most extreme Eigenmood axis (EM2, positive direction).

Reference	Score	English translation
Melancholia	$0.950$	At every moment I weep; I weep from the grief of time. (Translation ours.)
Melancholia	$0.950$	So long as Iraqi remains in the pain of parting, speak not of cure and do not try to heal me. (Translation ours.)
Melancholia	$0.950$	I plucked no flower from the garden of joy; from grief, a hundred thorns were lodged in my soul. (Translation ours.)

Appendix

Table B35: Eraghi: top verses for the leading concept (Melancholia).

Poet	Verses	Abstain	$D_{JS}$
Shahriar	$1970$	$0.141$	$0.0030$

Appendix

Table B36: Core profile metrics for Shahriar .

Poet	EM1	EM2	EM3
Shahriar	$-0.007$	$0.010$	$-0.012$

Appendix

Table B37: Eigenmood coordinates for Shahriar .

Rank	Concept	$P_{i}(c)$
1	Melancholia	$0.348$
2	Romantic Obsession	$0.175$
3	Emotional Dependency	$0.151$

Appendix

Table B38: Top concept shares for Shahriar . Concept names are abbreviated in the header; full interpretation appears in the text.

Figure B22: Shahriar: concept distribution

P_{i}(c)

against the global baseline

P_{0}(c)

Figure B23: Shahriar: lift profile

\Delta_{i}(c)=P_{i}(c)-P_{0}(c)

with positive and negative deviations.

Figure B24: Shahriar: bootstrap 95% intervals for the top-3 concepts by

P_{i}(c)

Reference	Score	English translation
EM3 (negative)	$-0.759$	If your heart would open, speak the truth of companionship: you were never my friend; I was yours. (Translation ours.)
EM3 (negative)	$-0.634$	Let the lamp of wakefulness cast its light; every heart’s eye is asleep in heedlessness, my friend. (Translation ours.)
EM3 (negative)	$-0.601$	The gates of gnosis are shut before your reason; the key to our coffer is love. (Translation ours.)

Appendix

Table B39: Shahriar: top verses on the poet’s most extreme Eigenmood axis (EM3, negative direction).

Reference	Score	English translation
Melancholia	$0.900$	What a world, if there is a heart and no heart-soothing; what a life, if grief is present and no consoler. (Translation ours.)
Melancholia	$0.850$	The candle of delight bloomed in tears and sighs; cloud and moon rose together. (Translation ours.)
Melancholia	$0.850$	No one has read your lot as joy beneath this turning sky; here, rejection and denial are the standing decree. (Translation ours.)

Appendix

Table B40: Shahriar: top verses for the leading concept (Melancholia).

Poet	Verses	Abstain	$D_{JS}$
Athir	$1603$	$0.138$	$0.0073$

Appendix

Table B41: Core profile metrics for Athir .

Poet	EM1	EM2	EM3
Athir	$-0.044$	$-0.007$	$0.004$

Appendix

Table B42: Eigenmood coordinates for Athir .

Rank	Concept	$P_{i}(c)$
1	Melancholia	$0.290$
2	Romantic Obsession	$0.222$
3	Emotional Dependency	$0.167$

Appendix

Table B43: Top concept shares for Athir . Concept names are abbreviated in the header; full interpretation appears in the text.

Figure B25: Athir: concept distribution

P_{i}(c)

against the global baseline

P_{0}(c)

Figure B26: Athir: lift profile

\Delta_{i}(c)=P_{i}(c)-P_{0}(c)

with positive and negative deviations.

Figure B27: Athir: bootstrap 95% intervals for the top-3 concepts by

P_{i}(c)

Reference	Score	English translation
EM1 (negative)	$-0.327$	How can one carry away the heart from my house, when the hand that steals it rises from within? (Translation ours.)
EM1 (negative)	$-0.275$	I saw that all accounts are zero beside the tally of your pain; I inscribed that tally upon myself. (Translation ours.)
EM1 (negative)	$-0.275$	Your grief plundered my soul, body, faith, and heart; keep your evil eye away, master thief that you are. (Translation ours.)

Appendix

Table B44: Athir: top verses on the poet’s most extreme Eigenmood axis (EM1, negative direction).

Reference	Score	English translation
Melancholia	$0.900$	Who sits with my heart but grief? Who keeps company with my soul but your pain? (Translation ours.)
Melancholia	$0.850$	Little of the night of sorrow remains; we must weep farewell to the day of joy. (Translation ours.)
Melancholia	$0.850$	We are left with a life held in grief’s hand: much of life is gone, and little patience remains. (Translation ours.)

Appendix

Table B45: Athir: top verses for the leading concept (Melancholia).

Poet	Verses	Abstain	$D_{JS}$
Khayyam	$356$	$0.385$	$0.0901$

Appendix

Table B46: Core profile metrics for Khayyam .

Poet	EM1	EM2	EM3
Khayyam	$0.024$	$-0.059$	$-0.099$

Appendix

Table B47: Eigenmood coordinates for Khayyam .

Rank	Concept	$P_{i}(c)$
1	Melancholia	$0.538$
2	Identity Fragmentation	$0.147$
3	Spiritual Narcissism	$0.101$

Appendix

Table B48: Top concept shares for Khayyam . Concept names are abbreviated in the header; full interpretation appears in the text.

Figure B28: Khayyam: concept distribution

P_{i}(c)

against the global baseline

P_{0}(c)

Figure B29: Khayyam: lift profile

\Delta_{i}(c)=P_{i}(c)-P_{0}(c)

with positive and negative deviations.

Figure B30: Khayyam: bootstrap 95% intervals for the top-3 concepts by

P_{i}(c)

Reference	Score	English translation
EM3 (negative)	$-0.724$	The shell hides that drop in which the sea-heart’s secret resides. (Translation ours.)
EM3 (negative)	$-0.637$	We are the source of joy and also the mine of sorrow. (Translation ours.)
EM3 (negative)	$-0.601$	We are lowliness and loftiness, fullness and lack; a tarnished mirror and Jamshid’s cup. (Translation ours.)

Appendix

Table B49: Khayyam: top verses on the poet’s most extreme Eigenmood axis (EM3, negative direction).

Reference	Score	English translation
Melancholia	$0.850$	My life has long been dark and my work unstraight; hardship has increased and ease has grown scarce. (Translation ours.)
Melancholia	$0.850$	That crystal cup that laughs with wine is a tear concealing the heart’s blood. (Translation ours.)
Melancholia	$0.850$	Drink wine, for life is one death pursues; better it pass in sleep or in intoxication. (Translation ours.)

Appendix

Table B50: Khayyam: top verses for the leading concept (Melancholia).

Spectral Graph Embedding

See all 16 papers

Mar 11, 2026

Artificial Intelligence for Sentiment Analysis of Persian Poetry

Recent advancements of the Artificial Intelligence (AI) have led to the development of large language models (LLMs) that are capable of understanding, analysing, and creating textual data. These language models open a significant opportunity in analyzing the literature and more specifically poetry. In the present work, we employ multiple Bidirectional encoder representations from transformers (BERT) and Generative Pre-trained Transformer (GPT) based language models to analyze the works of two prominent Persian poets: Jalal al-Din Muhammad Rumi (Rumi) and Parvin E'tesami. The main objective of this research is to investigate the capability of the modern language models in grasping complexities of the Persian poetry and explore potential correlations between the poems' sentiment and their meters. Our findings in this study indicates that GPT4o language model can reliably be used in analysis of Persian poetry. Furthermore, the results of our sentiment analysis revealed that in general, Rumi's poems express happier sentiments compared to Parvin E'tesami's poems. Furthermore, comparing the utilization of poetic meters highlighted Rumi's poems superiority in using meters to express a wider variety of sentiments. These findings are significant as they confirm that LLMs can be effectively applied in conducting computer-based semantic studies, where human interpretations are not required, and thereby significantly reducing potential biases in the analysis.

Arash Zargar, Abolfazl Moshiri, Mitra Shafaei +3

Department of Mechanical & Industrial Engineering; Vector Institute for Artificial Intelligence; University of Toronto; Toronto, Canada · Department of Near & Middle Eastern Civilizations; University of Toronto; Toronto, Canada · Independent Researcher +1

Computational Literary Studies Sentiment Analysis

Ask PDF

Apr 8, 2026

Between Century and Poet: Graph-Based Lexical Semantic Change in Persian Poetry

Meaning in Persian poetry is both historical and relational. Words persist through literary tradition while shifting their force through changing constellations of neighbors, rhetorical frames, and poetic voices. This study examines that process using aligned Word2Vec spaces combined with graph-based neighborhood analysis across centuries and major poets. Rather than modeling semantic change as vector displacement alone, it treats lexical history as the rewiring of local semantic graphs: the gain and loss of neighbors, shifts in bridge roles, and movement across communities. The analysis centers on twenty target words, anchored by five recurrent reference terms: Earth, Night, two wine terms, and Heart. Surrounding them are affective, courtly, elemental, and Sufi concepts such as Love, Sorrow, Dervish, King, Annihilation, and Truth. These words exhibit distinct patterns of change. Night is more time-sensitive, Earth more poet-sensitive, and Heart shows continuity despite graph-role mobility. The two wine terms highlight probe sensitivity: one is broad and semantically diffuse, while the other is narrower and more stable. A lexical audit confirms that the corpus contains historically driven terms, poet-specific usages, and sparsely attested mystical vocabulary requiring caution. Overall, semantic change in Persian poetry is better captured as neighborhood rewiring than as abstract drift. For Digital Humanities, this approach restores local structure to computational analysis and supports interpretations closer to literary practice: persistence, migration, mediation, and selective transformation.

Kourosh Shahnazari, Seyed Moein Ayyoubzadeh, Mohammadali Keshtparvar

Sharif University of Technology · Amirkabir University of Technology

Semantic Change Detection Computational Literary Studies

Ask PDF

Mar 15, 2026

Echoes Across Centuries: Phonetic Signatures of Persian Poets

This study examines phonetic texture in Persian poetry as a literary-historical phenomenon rather than a by-product of meter or a feature used only for classification. The analysis draws on a large corpus of 1,116,306 mesras from 31,988 poems written by 83 poets, restricted to five major classical meters to enable controlled comparison. Each line is converted into a grapheme-to-phoneme representation and analyzed using six phonetic metrics: hardness, sonority, sibilance, vowel ratio, phoneme entropy, and consonant-cluster ratio. Statistical models estimate poet-level differences while controlling for meter, poetic form, and line length. The results show that although meter and form explain a substantial portion of phonetic variation, they do not eliminate systematic differences between poets. Persian poetic sound therefore appears as conditioned variation within shared prosodic structures rather than as either purely individual style or simple metrical residue. A multidimensional stylistic map reveals several recurrent phonetic profiles, including high-sonority lyric styles, hardness-driven rhetorical or epic styles, sibilant mystical contours, and high-entropy complex textures. Historical analysis indicates that phonetic distributions shift across centuries, reflecting changes in genre prominence, literary institutions, and performance contexts rather than abrupt stylistic breaks. The study establishes a corpus-scale framework for phonetic analysis in Persian poetry and demonstrates how computational phonetics can contribute to literary-historical interpretation while remaining attentive to the formal structures that shape Persian verse.

Kourosh Shahnazari, Seyed Moein Ayyoubzadeh, Mohammadali Keshtparvar

Sharif University of Technology, Tehran, Iran. · Amirkabir University of Technology, Tehran, Iran.

Topic pending

Ask PDF

Abstract

Spectral Graph Embedding

Explore similar work

Artificial Intelligence for Sentiment Analysis of Persian Poetry

Between Century and Poet: Graph-Based Lexical Semantic Change in Persian Poetry

Echoes Across Centuries: Phonetic Signatures of Persian Poets