From Promises to Accountability: Comparative AI Modeling and Visualization of Electoral Commitments in Nepali Political Texts

Prakash Poudyal; Nischal Karki; Amrit Dahal; Rajani Chulyadyo; Sameer Tamrakar; Bal Krishna Bal

cs.CLOct 23, 2026

From Promises to Accountability: Comparative AI Modeling and Visualization of Electoral Commitments in Nepali Political Texts

Authors: Prakash Poudyal, Nischal Karki, Amrit Dahal, Rajani Chulyadyo, Sameer Tamrakar, Bal Krishna Bal

Organizations: Information and Language Processing Research Lab, Kathmandu University, Kavre, Nepal · Department of Artificial Intelligence, Kathmandu University, Nepal · Department of Computer Science and Engineering, Kathmandu University, Nepal

Abstract

Election Commitment Mining refers to the analysis of election commitments made by electoral candidates or political parties during election campaigns. However, many elected candidates and political parties, especially in developing South Asian countries, do not fully deliver on the promises they make before elections. Even after the elections, commitments presented through election manifestos, posters, pamphlets, videos, and other campaign materials are often not systematically or transparently archived. As a result, these promises tend to fade away over time, making it difficult to hold the elected candidates accountable for their commitments. To address this persistent issue, this research aims to develop a system capable of structuring the commitments made by candidates during election periods. In this work, we created the first gold-standard dataset of Nepali political texts (Cohen's kappa = 0.777) and evaluated machine learning, deep learning and large language models to automatically extract and analyze election commitments. LaBSE-LR achieved the best results among machine learning models (0.897 accuracy, 0.896 macro F1-score, 0.947 AUROC), while fastText-LSTM achieved the highest performance among deep learning models (0.914 accuracy, 0.913 macro F1-score, 0.961 AUROC). Among the three evaluated large language models, Qwen3-8B performed best in the few-shot setting (0.848 macro F1-score). The outcome of this work is integrated into a mobile application, where commitments can be visualized with the support of GAI, enhancing transparency, civic engagement, and electoral accountability.

Figures & tables

Figure 1: System architecture for mining commitments from Nepali political texts

Provinces	Nagarpalikas	People	Images	Videos
Koshi	8	8	32	6
Madhesh	5	6	27	2
Bagmati	13	19	28	12
Lumbini	9	10	18	9
Sudurpashchim	2	2	14	0
Total	37	45	119	29

Table 1: Dataset composition by province: candidates, images, and videos collected from Facebook profiles and party websites

Figure 2: Preprocessing pipeline for pamphlet image-to-text conversion: from Google Lens and Tesseract OCR to refined data

		Annotator 1 (A1)
	Label	C	NC	\engbold $N_{A1}$
Annotator 2 (A2)	C	1,254	223	1,477
	NC	44	909	953
	\engbold $N_{A2}$	1,298	1,132	$N=2{,}430$

Table 2: Inter-annotator agreement confusion matrix

Provinces	No. of Candidates/Parties	No. of Commitments (C)	No. of Non-Commitments (NC)
Koshi	6	228	99
Madhesh	6	196	64
Bagmati	11	688	899
Lumbini	3	108	30
Sudurpashchim	2	101	17
Total	28	1,321	1,109

Table 3: Distribution of commitment and non-commitment statements by province

Panel of Figure 3 — Figure 3: Profile images of the candidates as displayed in the mobile application

Figure 4: Workflow diagram of the mobile application

Name of Candidate	Commitments (in Nepali)	Summarized Commitments (in Nepali)	Commitments (in English)
\devanagarifont बालेन्द्र शाह [Figure 3(a) ]	\devanagarifont काठमाडौं महानगरपालिकाका सामुदायिक विद्यालयहरूको पूर्वाधार, गुणस्तर र वातावरणमा सुधार गर्नु	\devanagarifont सामुदायिक विद्यालयको गुणस्तर सुधार	To improve the infrastructure, quality, and environment of community schools in Kathmandu Metropolitan City and to emphasize technical education.
	\devanagarifont अन्तर्राष्ट्रिय विद्यालयहरूसँगको सहकार्य गरी जीवनशैली, संस्कृति, शिक्षा प्रणाली आदिको आदानप्रदान गर्नु	\devanagarifont अन्तर्राष्ट्रिय विद्यालयसँग जीवनशैलीको आदानप्रदान	International schools will be partnered with national schools and sisterhood relationships will be initiated to exchange culture, lifestyle, education systems, etc.
\devanagarifont सुरेश साह सोनार [Figure 3(b) ]	\devanagarifont कृषि क्षेत्रको प्रवर्द्धन एवं विकासको लागि किसानलाई सहज तथा समयमा सहुलियत रकममा बिउ, बिजन र कृषि उपकरण उपलब्ध गराउने छु	\devanagarifont किसानलाई सहुलियतमा बिउ र उपकरण उपलब्ध गराउने	For the promotion and development of the agricultural sector, I will provide seeds, seedlings and agricultural equipment to farmers easily and on time at subsidized rates.
	\devanagarifont प्रत्येक सामुदायिक विद्यालयमा अंग्रेजी माध्यमबाट पठन पाठन संचालन गर्नको साथै उच्च-शिक्षा हासिल गर्न जेहेन्दार छात्र छात्राहरुलाई छात्रवृत्ति प्रदान गर्ने छु	\devanagarifont अंग्रेजी माध्यममा जेहेन्दार विद्यार्थीलाई छात्रवृत्ति दिनेछ	I will provide scholarships to meritorious students to pursue higher education by conducting English medium teaching in every community school.

Table 4: A few example candidate commitments shown in original Nepali, summarized Nepali, and English translation

Model Name	Accuracy	ma-Precision	ma-Recall	ma- $F_{1}$ -score	AUROC
TF-IDF-LR	0.870	0.869	0.869	0.869	0.943
LaBSE-LR	\engbold 0.897	\engbold 0.896	\engbold 0.896	\engbold 0.896	0.947
TF-IDF-DT	0.735	0.736	0.740	0.735	0.809
LaBSE-DT	0.747	0.744	0.741	0.742	0.813
TF-IDF-RF	0.860	0.858	0.861	0.859	0.931
LaBSE-RF	0.870	0.871	0.866	0.868	0.939

Table 5: Performance of ML models under TF-IDF and LaBSE text representations. The best result is \engbold bolded, and the second best is underlined .

Model Name	Accuracy	ma-Precision	ma-Recall	ma- $F_{1}$ -score	AUROC
fastText-SRN	0.899	0.898	0.899	0.898	0.951
GloVe-SRN	0.852	0.850	0.854	0.851	0.924
fastText-LSTM	\engbold 0.914	\engbold 0.912	\engbold 0.915	\engbold 0.913	\engbold 0.961
GloVe-LSTM	0.840	0.841	0.845	0.839	0.915
fastText-GRU	0.905	0.904	0.904	0.904	0.959
GloVe-GRU	0.850	0.849	0.853	0.849	0.920

Table 6: Performance of DL models under fastText and GloVe embeddings, and transformer-based models. The best result is \engbold bolded, and the second best is underlined .

Model Name	Accuracy	ma-Precision	ma-Recall	ma- $F_{1}$ -score	AUROC
Gemma-4-E4B-it- \engital zero-shot	0.815	0.819	0.822	0.815	0.822
Gemma-4-E4B-it- \engital few-shot	0.833	0.838	0.841	0.833	0.841
Llama-3.1-8B-Instruct- \engital zero-shot	0.609	0.687	0.639	0.593	0.639
Llama-3.1-8B-Instruct- \engital few-shot	0.842	0.844	0.848	0.841	0.848
Qwen3-8B- \engital zero-shot	0.817	0.831	0.829	0.817	0.829
Qwen3-8B- \engital few-shot	\engbold 0.848	\engbold 0.855	\engbold 0.857	\engbold 0.848	\engbold 0.857

Table 7: Performance of LLMs under zero-shot and few-shot prompting. The best result is \engbold bolded, and the second best is underlined .

Model A	Model B	$b$	$c$	$n$	$p_{\text{mid-}p}$
fastText-SRN	fastText-LSTM	10	16	26	0.248
fastText-GRU	fastText-LSTM	12	15	27	0.572
\engbold fastText-CNN	\engbold fastText-LSTM	\engbold 6	\engbold 17	\engbold 23	\engbold 0.0227
\engbold XLM-R	\engbold fastText-LSTM	\engbold 10	\engbold 24	\engbold 34	\engbold 0.0167
\engbold TF-IDF-LR	\engbold fastText-LSTM	\engbold 10	\engbold 30	\engbold 40	\engbold 0.00145
\engbold TF-IDF-SVM	\engbold fastText-LSTM	\engbold 10	\engbold 30	\engbold 40	\engbold 0.00145

Table 8: Pairwise McNemar mid-

p

test results. Model B denotes the higher-performing model (higher

F_{1}

-score) in each comparison.

b

= pairs where Model A is correct and Model B is wrong;

c

= pairs where Model A is wrong and Model B is correct;

n=b+c

. The statistically significant pairs are \engbold bolded.

Model	Accuracy	ma-Precision	ma-Recall	ma- $F_{1}$ -score	AUROC
TF-IDF-LR \engital w augmentation	0.870	0.869	0.869	0.869	0.943
TF-IDF-LR \engital w/o augmentation	0.848	0.846	0.846	0.846	0.938
$\Delta$ TF-IDF-LR	\engbold +2.20%	\engbold +2.30%	\engbold +2.30%	\engbold +2.30%	\engbold +0.50%
LaBSE-LR \engital w augmentation	0.897	0.896	0.896	0.896	0.947
LaBSE-LR \engital w/o augmentation	0.897	0.896	0.896	0.896	0.947
$\Delta$ LaBSE-LR	0.00%	0.00%	0.00%	0.00%	0.00%

Table 9: Ablation Study 1: Impact of data augmentation across embedding families and architectures. \engital w augmentation denotes training with 155 synthetic-reviewed NC instances added to the training set, \engital w/o augmentation denotes training on original data only.

\Delta

indicates change over the no-augmentation baseline. Meaningful improvement are \engbold bolded (

\geq 1.00\%

Model	Accuracy	ma-Precision	ma-Recall	ma- $F_{1}$ -score	AUROC
fastText-LSTM \engital (fine-tuned)	0.914	0.912	0.915	0.913	0.961
fastText-LSTM \engital (pre-trained)	0.850	0.848	0.852	0.849	0.922
$\Delta$ fastText-LSTM	\engbold +6.40%	\engbold +6.40%	\engbold +6.30%	\engbold +6.40%	\engbold +3.90%
fastText-GRU \engital (fine-tuned)	0.905	0.904	0.904	0.904	0.959
fastText-GRU \engital (pre-trained)	0.852	0.850	0.854	0.851	0.934
$\Delta$ fastText-GRU	\engbold +5.30%	\engbold +5.40%	\engbold +5.00%	\engbold +5.30%	\engbold +2.50%

Table 10: Ablation Study 2: Impact of fine-tuning pretrained embeddings on the electoral commitment dataset. \engital fine-tuned denotes pretrained embeddings further trained on the electoral commitment training corpus, \engital pre-trained denotes frozen pretrained embeddings used without any domain adaptation.

\Delta

indicates change of fine-tuned over pre-trained baseline.Meaningful improvement are \engbold bolded (

\geq 1.00\%

Panel of Figure 5 — Figure 5: Confusion matrices for LaBSE-LR, fastText-LSTM, XLM-RoBERTa, and Qwen3-8B (few-shot).

Factor	Statement (ne)	Statement (en)	Error Type
	\devanagarifont युवाहरुलाई सीपमूलक तालिमको व्यवस्था गरी लघु तथा घरेलु उद्योग सञ्चालनमा प्रोत्साहन गरिने छ।	Youth will be encouraged to operate small and cottage industries by providing skill-based training.	FP
Lexical cue	\devanagarifont प्रत्येक तहमा जिम्मेवारी विने, कामको मुल्याङ्‌कन गर्ने र पदोन्नति गर्ने वैज्ञानिक पद्धतिको निर्माण गरी अवलम्वन गरिने छ।	A scientific system for assigning responsibilities, evaluating work, and promoting will be developed and adopted at each level.	FP
	\devanagarifont नगरभित्र कवड हल र रतुवामाई ४ गढीमाई महाराजथानमा रंगशाला निर्माणका लागि आवश्यक पहल गरिने छ।	Necessary initiatives will be taken to construct a stadium in the city’s Kawad Hall and Ratuwamai 4 Gadhimai Maharajthan.	FP
	\devanagarifont ईन्टरनेट सेवा जिल्ला भर	Internet service throughout the district	FN
	\devanagarifont उदाहरणीय सुशासन	Exemplary Good Governance	FN
Syntactic completeness	\devanagarifont जस अन्तरगत ५ वर्ष भित्र १ लाख नयाँ बोट रोपनि छन	Under which 1 lakh new trees will be planted within 5 years.	FN

Table 11: Representative example error cases grouped by linguistic factor. FP = false positive (NC predicted as C), FN = false negative (C predicted as NC).

Appendix figures & tables2 assets

Supplementary material from the paper’s appendix.

Appendix

Model	Text Representation	Hyperparameters
LR	TF-IDF	C=10.0, class_weight=None, penalty=l2, max_features=20000, ngram_range=(1, 2)
LR	LaBSE	C=1.0, class_weight=balanced, penalty=l2
DT	TF-IDF	class_weight=None, criterion=entropy, max_features=5000, ngram_range=(1, 1), max_depth=20, min_samples_split=10, min_samples_leaf=2
DT	LaBSE	class_weight=balanced, criterion=entropy, max_depth=5, min_samples_split=2, min_samples_leaf=1
RF	TF-IDF	class_weight=None, n_estimators=300, max_features=5000, max_depth=None, min_samples_split=2, min_samples_leaf=1, ngram_range=(1, 2)
RF	LaBSE	class_weight=None, n_estimators=100, max_depth=10, min_samples_split=10, min_samples_leaf=2

Appendix

Table 12: Optimal hyperparameter settings for ML models under TF-IDF and LaBSE text representations

Model	Text Representation	Model Architecture	Hyperparameters
SRN	fastText	Embedding (max_length=64, dim=300), SRN (36 units), Sigmoid (1)	Batch size=16, epochs=30, patience=3, lr=3e-4, dropout=0.3, l2_reg=0.0005
SRN	GloVe	Embedding (max_length=64, dim=300), SRN (48 units), Dense (32 units), Sigmoid (1)	Batch size=16, epochs=30, patience=3, lr=3e-4, dropout=0.5, l2_reg=0.001
LSTM	fastText	Embedding (max_length=64, dim=300), LSTM (36 units), Sigmoid (1)	Batch size=16, epochs=30, patience=3, lr=3e-4, dropout=0.3, l2_reg=0.0005
LSTM	GloVe	Embedding (max_length=64, dim=300), LSTM (48 units), Dense (24 units), Sigmoid (1)	Batch size=16, epochs=30, patience=3, lr=5e-4, dropout=0.3, l2_reg=0.0005
GRU	fastText	Embedding (max_length=64, dim=300), GRU (36 units), Sigmoid (1)	Batch size=16, epochs=30, patience=3, lr=3e-4, dropout=0.3, l2_reg=0.0005
GRU	GloVe	Embedding (max_length=64, dim=300), GRU (36 units), Dense (24 units), Sigmoid (1)	Batch size=16, epochs=30, patience=3, lr=5e-4, dropout=0.3, l2_reg=0.001

Appendix

Table 13: Architecture and optimal hyperparameter settings for DL models under fastText, GloVe, and transformer-based representations

Seeing Candidates at Scale: Multimodal LLMs for Visual Political Communication on Instagram
Apr 21, 2026Michael Achmann-Denkler, Mario Haim, Christian WolffVLM Evaluation Social Media Analysis
Media Informatics Group University of Regensburg Regensburg, Germany · Department of Media and Communication Ludwig-Maximilians-Universität Munich, Germany
Who Would You Vote For? Auditing Political Alignment in LLMs: An Italian Case-Study
Aug 12, 2026Simone MungariLLM Auditing Political Bias in Language Models
Revelis s.r.l.
Large Language Models as Implicit Sociological Models: Reconstructing Voting Behaviour from Sociodemographic Profiles
Aug 16, 2026Roman Neruda, Martin Bakoš, Josef Šlerka +3Computational Social Science Large Language Model-Based Social Simulation
Institute of Computer Science, Czech Academy of Sciences, Prague, Czech Republic. · Faculty of Arts, Charles University, Prague, Czech Republic. · distil labs, Berlin, Germany.

Apr 21, 2026J/K move · Enter open · S save

Seeing Candidates at Scale: Multimodal LLMs for Visual Political Communication on Instagram

Michael Achmann-Denkler, Mario Haim, Christian Wolff

Media Informatics Group University of Regensburg Regensburg, Germany · Department of Media and Communication Ludwig-Maximilians-Universität Munich, Germany

VLM Evaluation Social Media Analysis Face Recognition

Open paper AskPDF arXiv

This paper presents a computational case study that evaluates the capabilities of specialized machine learning models and emerging multimodal large language models for Visual Political Communication (VPC) analysis. Focusing on concentrated visibility in Instagram stories and posts during the 2021 German federal election campaign, we compare the performance of traditional computer vision models (FaceNet512, RetinaFace, Google Cloud Vision) with a multimodal large language model (GPT-4o) in identifying front-runner politicians and counting individuals in images. GPT-4o outperformed the other models, achieving a macro F1-score of 0.89 for face recognition and 0.86 for person counting in stories. These findings demonstrate the potential of advanced AI systems to scale and refine visual content analysis in political communication while highlighting methodological considerations for future research.

Abstract

Explore similar work

Seeing Candidates at Scale: Multimodal LLMs for Visual Political Communication on Instagram

Who Would You Vote For? Auditing Political Alignment in LLMs: An Italian Case-Study

Large Language Models as Implicit Sociological Models: Reconstructing Voting Behaviour from Sociodemographic Profiles