Paper deep dive
Evaluating Cross-Lingual Classification Approaches Enabling Topic Discovery for Multilingual Social Media Data
Deepak Uniyal, Md Abul Bashar, Richi Nayak
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 94%
Last extracted: 7/21/2026, 12:37:44 AM
Summary
This study evaluates four cross-lingual text classification approaches for filtering relevant hydrogen energy tweets from a decade-long, multilingual dataset (English, Japanese, Hindi, Korean) collected via keyword search. The approaches include language-specific models, an English-only translation-based model, a zero-shot multilingual BERT (mBERT) model, and a hybrid mBERT model. The English-only translation approach yielded the best performance for relevance classification, enabling effective topic modeling using Non-negative Matrix Factorization (NMF) to uncover thematic trends in hydrogen energy discourse from 2013 to 2022.
Entities (14)
Relation Signals (15)
Deepak Uniyal â coauthorof â Evaluating Cross-Lingual Classification Approaches Enabling Topic Discovery for Multilingual Social Media Data
confidence 98% · Evaluating Cross-Lingual Classification approaches Enabling Topic Discovery for Multilingual Social Media Data Deepak Uniyal
Md Abul Bashar â coauthorof â Evaluating Cross-Lingual Classification Approaches Enabling Topic Discovery for Multilingual Social Media Data
confidence 98% · Deepak Uniyal ... Md Abul Bashar ... and Richi Nayak
Richi Nayak â coauthorof â Evaluating Cross-Lingual Classification Approaches Enabling Topic Discovery for Multilingual Social Media Data
confidence 98% · Deepak Uniyal ... Md Abul Bashar ... and Richi Nayak
Deepak Uniyal â affiliatedwith â Queensland University of Technology
confidence 95% · Deepak Uniyal ... Centre for Data Science, School of Computer Science, Queensland University of Technology
Md Abul Bashar â affiliatedwith â Queensland University of Technology
confidence 95% · Md Abul Bashar ... Centre for Data Science, School of Computer Science, Queensland University of Technology
Richi Nayak â affiliatedwith â Queensland University of Technology
confidence 95% · Richi Nayak ... Centre for Data Science, School of Computer Science, Queensland University of Technology
Evaluating Cross-Lingual Classification Approaches Enabling Topic Discovery for Multilingual Social Media Data â analyzes â Hydrogen Energy
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Analysing multilingual social media discourse remains a major challenge in natural language processing, particularly when large-scale public debates span across diverse languages. This study investigates how different approaches for cross-lingual text classification can support reliable analysis of global conversations. Using hydrogen energy as a case study, we analyse a decade-long dataset of over nine million tweets in English, Japanese, Hindi, and Korean (2013--2022) for topic discovery. The online keyword-driven data collection results in a significant amount of irrelevant content. We explore four approaches to filter relevant content: (1) translating English annotated data into target languages for building language-specific models for each target language, (2) translating unlabelled data appearing from all languages into English for creating a single model based on English annotations, (3) applying English fine-tuned multilingual transformers directly to each target language data, and (4) a hybrid strategy that combines translated annotations with multilingual training. Each approach is evaluated for its ability to filter hydrogen-related tweets from noisy keyword-based collections. Subsequently, topic modeling is performed to extract dominant themes within the relevant subsets. The results highlight key trade-offs between translation and multilingual approaches, offering actionable insights into optimising cross-lingual pipelines for large-scale social media analysis.
Tags
Links
- Source: https://arxiv.org/abs/2602.17051v1
- Canonical: https://arxiv.org/abs/2602.17051v1
Trouble viewing inline? Open PDF directly â
Full Text
41,280 characters extracted from source content.
Expand or collapse full text
Evaluating Cross-Lingual Classification approaches Enabling Topic Discovery for Multilingual Social Media Data Deepak Uniyal 1[0000â0002â7665â1585] , Md Abul Bashar 2[0000â0003â1004â4085] , and Richi Nayak 3[0000â0002â9954â0159] Centre for Data Science, School of Computer Science, Queensland University of Technology, Brisbane, QLD 4000, Australia 1 deepak.uniyal@hdr.qut.edu.au, 2 m1.bashar@qut.edu.au, 3 r.nayak@qut.edu.au AbstractAnalysing multilingual social media discourse remains a ma- jor challenge in natural language processing, particularly when large- scale public debates span across diverse languages. This study inves- tigates how different approaches for cross-lingual text classification can support reliable analysis of global conversations. Using hydrogen energy as a case study, we analyse a decade-long dataset of over nine million tweets in English, Japanese, Hindi, and Korean (2013â2022) for topic discovery. The online keyword-driven data collection results in a signif- icant amount of irrelevant content. We explore four approaches to filter relevant content: (1) translating English annotated data into target lan- guages for building language-specific models for each target language, (2) translating unlabelled data appearing from all languages into English for creating a single model based on English annotations, (3) apply- ing English fine-tuned multilingual transformers directly to each target language data, and (4) a hybrid strategy that combines translated an- notations with multilingual training. Each approach is evaluated for its ability to filter hydrogen-related tweets from noisy keyword-based col- lections. Subsequently, topic modeling is performed to extract dominant themes within the relevant subsets. The results highlight key trade-offs between translation and multilingual approaches, offering actionable in- sights into optimising cross-lingual pipelines for large-scale social media analysis. Keywords:multilingual data · text classification · topic modelling · hydrogen energy · cross-lingual 1 Introduction The growth of social media platforms such as Twitter 1 has enabled the large-scale collection of user-generated content across multiple languages, offering opportu- nities for studying global public opinion [ 1]. A key challenge in this domain is 1 Recently changed its name toX. In this paper, we refer to this platform as Twitter. arXiv:2602.17051v1 [cs.CL] 19 Feb 2026 2Deepak Uniyal, Md Abul Bashar, and Richi Nayak relevance classification, i.e. identifying tweets pertinent to a specific topic across diverse linguistic contexts. The multilingual nature of such data introduces sig- nificant challenges for natural language processing (NLP), including handling imbalances in language resources, preserving cultural nuances, and mitigating translation-induced noise [2]. These issues fuel an ongoing debate: Should mul- tilingual social media data be processed using separate native language-specific models, translated into a pivot language (e.g., English), or directly handled by multilingual models such as multilingual BERT (mBERT) [3] and XLM- RoBERTa (XLM-R) [4] [5] [6]. While NLP research has advanced multilingual text processing methods, the optimal strategy remains context-dependent. Translating all data to a pivot language like English [ 7] simplifies downstream tasks but risks losing language- specific subtleties. Training separate language-specific models can preserve these nuances, yet it demands extensive annotated data for each language, which is often unavailable. Multilingual transformer models, such as mBERT [3] and XLM-R [4], offer the ability to process multiple languages directly; however, their performance varies depending on the training resources available for specific languages, often declining for low-resource language pairs [8]. To address these challenges, this study systematically evaluates four cross- lingual relevance classification approaches for large-scale, multi-year Twitter data. The dataset for this study comprises a large corpus of Twitter data span- ning multiple years, collected in English, Japanese, Korean, and Hindi. How- ever, a critical constraint is the limited availability of annotated data, a common challenge in real-world scenarios due to resource limitations. Only a small set of 5,000 English tweets is annotated for relevance, while the vast majority of the data, comprising millions of unlabelled tweets across all four languages remains unannotated. This scarcity of labeled data complicates the development of ro- bust cross-lingual systems, necessitating innovative approaches to leverage the available annotations effectively. The first approach employs language-specific modeling, translating the 5,000 English-annotated tweets into Japanese, Korean, and Hindi to train separate classifiers, while using the original English annotations for an English model. The second approach adopts a translation-based pipeline, translating all non- English tweets into English and applying a single English classifier. The third approach leverages mBERT, fine-tuned on the English annotations and applied in a zero-shot manner [ 9] to original-language tweets. Finally, the hybrid ap- proach combines translation and multilingual training, translating English anno- tations into multiple languages, merging them into a single multilingual training set, and fine-tuning mBERT for cross-lingual prediction. The main contributions of this study are as follows: 1.Cross-Lingual Classification. A comparative evaluation of four approaches for relevance classification in large-scale multilingual Twitter data, enhancing the filtering of hydrogen-energy content. 2.Key Themes and Phrases. An integrated pipeline using topic modeling to uncover thematic trends over a decade. Towards Effective Topic Discovery in Multilingual Data3 3.Temporal Evolution. Insights into the evolution of hydrogen energy discourse from 2013 to 2022, including periods of heightened user engagement. 4.First Decade-long Multilingual Analysis. A pioneering decade-long, multilin- gual analysis of hydrogen energy discourse across English, Japanese, Hindi, and Korean, with a scalable framework for future clean energy studies. The paper is organised as follows: Section 2covers related work, Section3 details our methodology, Section4presents results, and Section5offers conclu- sions and future directions. 2 Literature Review Cross-lingual NLP has progressed from early translation-based methods to ad- vanced multilingual models, addressing the challenge of linguistic diversity in multilingual datasets. The foundational âtranslate-then-classifyâ approach, pi- oneered by early work in the field, translates source text into a pivot language (e.g., English) to enable processing, though its effectiveness hinges on translation accuracy and struggles with semantic nuances [10]. Subsequent advancements in- troduced cross-lingual alignment techniques, such as structural correspondence learning, which identifies consistent pivot features across languages, reducing reliance on machine translation quality [11]. The development of multilingual word embeddings further enhanced semantic similarity across languages through linear transformations [ 12]. The advent of transformer-based models marked a significant leap, with Multilingual BERT (mBERT) demonstrating robust cross- lingual performance in tasks like classification and entity recognition [8]. En- hanced models like XLM and XLM-R, trained on larger corpora with refined objectives, have improved transfer capabilities across diverse languages [13] [14]. However, these models often underperform with low-resource languages. Recent hybrid approaches combine monolingual specialisation, which captures language- specific nuances, with multilingual generalisation for cross-lingual alignment, though their application in specialised domains remains underexplored. Our study addresses this gap by evaluating four cross-lingual relevance clas- sification approaches â monolingual, translation-based, multilingual, and hybrid â on a decade-long, four-language dataset, enabling effective filtering for topic modeling to uncover regional hydrogen energy themes. 3 Methodology This study presents a systematic pipeline for analysing multilingual Twitter dis- course on hydrogen energy spanning 2013 to 2022. The methodology, illustrated in Figure1, integrates data collection, preprocessing, and relevance classification, which enable further NLP tasks such as topic modeling or sentiment mining, based on the best method. The central challenge arises from the multilingual nature of the corpus: Tweets appear in English (EN), Japanese (JA), Korean 4Deepak Uniyal, Md Abul Bashar, and Richi Nayak Figure1: Architecture of the proposed multi-step methodology, including data collection, data preprocessing, and relevance classification. (KO), and Hindi (HI), while supervised annotations are initially available only in English. To address this challenge, we developed and evaluated four distinct cross- lingual relevance classification approaches: (1) Training independent language- specific models for each language, (2) Training a single English model on English annotated data, (3) Fine-tuning mBERT on English annotated data and employ- ing zero-shot predictions for other languages, and (4) Fine-tuning mBERT on English annotated data and its translation into other languages, followed by cross-lingual predictions. Following relevance classification using the optimal approach, topic modelling is performed on the filtered, thematically coherent dataset. 3.1 Data Collection and Preprocessing We collected a large-scale dataset of 30.7 million tweets related to hydrogen energy using the Academic Twitter API v2, covering the period from January 2013 to December 2022. Keywords were curated through a combination of lit- erature review and expert consultation to ensure comprehensive coverage of the discourse across four languages: English, Japanese, Hindi, and Korean [ 1]. The initial processing, as described in [1], included steps such as metadata mapping, deduplication, language-specific text normalisation, and tokenisation. Towards Effective Topic Discovery in Multilingual Data5 Table 1: Training and testing setup for four relevance-classification approaches. ApproachTraining DataTest DataModel Setup 1. Language- specific 5,000 English annotated tweets and their transla- tions (JA, KO, and HI) a 750 tweets each language taken from original language data Separate language-specific models for each language 2. English-only5,000 English annotated tweets Single English model; all tweets translated to English 3. English Fine- Tuned mBERT 5,000 English annotated tweets Single mBERT model fine- tuned on English data; zero- shot prediction on other non- English languages 4. Hybrid5,000 English annotated tweets and their transla- tions (JA, KO, and HI) a Single mBERT model trained on combined multilingual data a JA = Japanese, KO = Korean, HI = Hindi. After removing duplicate tweets, we retained approximately 9.9 million unique tweets. To ensure temporal consistency, we further filtered tweets strictly posted between 2013 and 2022. This step excluded retweets and other content origi- nating outside the target window but was re-shared during the period. The resulting dataset consisted of approximately 9.3 million tweets. For supervised relevance classification, a labeled dataset of 5,000 English tweets was created, uniformly sampled across years to maintain temporal cover- age. Each tweet was annotated asrelevantorirrelevantby three domain experts following detailed guidelines. Inter-annotator reliability, measured using Fleissâ kappa [15], achieved a score of 0.85, indicating strong agreement. Approximately 45% of tweets were labeled as irrelevant, highlighting the necessity of automated filtering. This annotated dataset provides the foundation for training all subse- quent relevance classification approaches. As we lack a labeled dataset in other languages, we translated these 5,000 tweets into Japanese, Korean, and Hindi usingfacebook/nllb-200-distilled- 600M[ 16] neural machine translation model. For a robust evaluation, we ran- domly selected and manually annotated 750 tweets each from original, non- translated Japanese, Hindi, and Korean tweets, ensuring uniform distribution across years and referencing their translation for accuracy. 3.2 Multilingual Relevance Classification The primary challenge in classifying the relevance of this dataset stems from its multilingual composition, with initial labels available only for the English subset. To address this, we developed and evaluated four distinct cross-lingual relevance classification approaches, which vary based on the training data composition and the type of encoder used (monolingual or multilingual), as detailed in Table 1. We first outline the experimental setup and data utilization across all approaches, followed by a detailed description of each approach: 6Deepak Uniyal, Md Abul Bashar, and Richi Nayak Experimental setupOur experimental framework incorporates a ten-run with ten different seeds, where the data for each language is divided into a train, validation and test sets in a ratio of 70%, 15%, and 15%, respectively. Depending on the chosen approach, a suitable subset of data is selected to train the model across ten seeds, and the best model is saved for prediction. We implement early stopping with a patience of 5 epochs, terminating training if no improvement in validation accuracy is observed. We evaluate the performance using accuracy and F1-score, as these metrics are widely used in text classification [17]. Both metrics are reported as mean and standard deviation over ten runs to ensure a robust evaluation. Since the original annotations were available only for English data, we cre- ated a uniform set of 750 annotated tweets per language (English, Japanese, Korean, and Hindi) from original language tweets, used exclusively for testing to evaluate model performance on authentic unseen tweets across all languages. These annotated samples are used to rigorously test each approach and report comparative performance metrics. Approach 1: Monolingual ModelsIn the first approach, we use 5,000 En- glish annotated tweets and their translations for all languages to fine-tune sep- arate monolingual classifiers independently. Each classifier is trained on its respective language data using a language-specific transformer model: 1.English:bert-base-uncased[3], a 12-layer bidirectional Transformer en- coder with 110M parameters, pre-trained on BooksCorpus and English Wikipedia. 2.Japanese:cl-tohoku/ bert-base-japanese[18], a BERT-base model pre- trained on Japanese Wikipedia, tokenised with a WordPiece-based vocabu- lary optimised for Japanese morphology. 3.Korean:monologg/kobert[19], a Korean BERT variant trained on a large- scale Korean corpus, adapted with a custom tokeniser to handle Korean morphology and spacing. 4.Hindi:ai4bharat/indic-bert[20], a multilingual BERT model covering 12 major Indian languages, trained on IndicCorp and designed for resource- limited Indic NLP tasks. For each model, the architecture consists of the language-specific encoder followed by a dropout layer (p=0.3), a fully connected projection layer, atanh activation, and a softmax output layer. Fine-tuning is performed independently for each language on corresponding labeled data. The performance of each model is tested on the uniform set of 750 annotated tweets per language from original language data. Approach 2: English-only ModelIn the second approach, we utilise anno- tated English tweets to train the English-only model, i.e.bert-base-uncased. The model architecture mirrors that of approach 1 ( 3.2) for prediction pur- poses. The test is performed on the English portion of 750 annotated tweets, Towards Effective Topic Discovery in Multilingual Data7 and additional tests are conducted on the English-translated versions of the 750 annotated tweets per language (Japanese, Korean, and Hindi) from the original language data. Approach 3: Multilingual ModelIn the third approach, we adopt a fully multilingual strategy by leveraging thebert-base-multilingual-uncased model, referred as mBERT [21], which supports 104 languages including English, Japanese, Korean, and Hindi. The fine-tuning phase utilises English annotated tweets to fine-tune the mBERT, relying on its generalisation capabilities from pre-training. The model architecture consists of the pre-trained mBERT encoder followed by a dropout layer withp=0.3, a fully connected projection layer, atanhactivation function, and a final softmax classification layer producing binary relevance predictions on the input data. The test is performed on the English portion of 750 annotated tweets per language, using the best model of ten runs for prediction. Additional tests are performed for non-English languages (Japanese, Korean, and Hindi) using the test set of 750 tweets in their native form. Approach 4: Hybrid Multilingual ModelThe fourth approach combines the benefits of the multilingual approach as in approach 3 (3.2), which relies on mBERT fine-tuned on English data with non-English translated data enriched with annotations, integrating these elements into a unified training framework. These translated annotations, together with the original English annotated data form a combined multilingual training set. This dataset is used to fine-tune an mBERT classifier with the same architecture as in approach 3. By using a mul- tilingual dataset for fine-tuning, this approach aims to improve the classification performance on non-English tweets while preserving the advantages of mBERTâs cross-lingual representation learning. The test is performed on the uniform set of 750 annotated tweets per language from original language data, and the best model from ten runs is saved for inference. 3.3 Topic Modelling Experimental results show that Approach 2 (the model trained with English- annotated tweets only) provided the most accurate relevancy distinction. After identifyingapproach 2as the best-performing and most interpretable strategy for relevance classification, we used its filtered output for topic modelling. One key advantage of this approach is that all tweets, regardless of their original lan- guage, are translated into English during the classification stage. This not only enables consistent processing but also makes the resulting themes more accessi- ble to a wider audience by presenting the discourse in a single language. Tweets labeled asirrelevantwere excluded from further analysis, ensuring that only content truly related to hydrogen energy shaped the thematic structure. We ex- plored topics from other approaches in their original languages, but their quality was noisy, likely due to challenges in handling multilingual data effectively. 8Deepak Uniyal, Md Abul Bashar, and Richi Nayak For topic discovery, we selected Non-negative Matrix Factorisation (NMF) [22] over methods like Latent Dirichlet Allocation (LDA) due to its interpretabil- ity, robustness, and superior alignment with human judgment [23]. NMF effec- tively derived coherent topics from sparse, high-dimensional TF-IDF represen- tations, which incorporated unigrams, bigrams, and trigrams, filtered infrequent terms, and limited the feature space to 10,000 terms for computational eïŹi- ciency. Configured to extract up to ten topics per language with reproducible initialisation, NMF ensured consistent comparability across datasets. This pipeline produced topic keywords with associated weights, tweet-to- topic assignments identifying each tweetâs dominant theme, and temporal distri- butions tracking yearly frequencies from 2013 to 2022. These outputs revealed the evolution of hydrogen energy discourse, highlighting periods of thematic growth or decline. By relying on a unified English-translated dataset, the ap- proach balanced model performance, interpretability, and comparability, mak- ing the analysis accessible to both technical and non-technical audiences while upholding rigorous topic extraction and trend analysis. 4 Results and Discussion This section provides a detailed analysis of the outcomes from the cross-lingual relevance classification and topic modeling applied to Twitter data on hydrogen energy discourse. The analysis focuses on temporal trends, including tweet dis- tribution over years and topic evolution, utilising the processed dataset derived fromapproach 2. The findings are interpreted in the context of hydrogen energy market trends, policy developments, and technological advancements, informed by prior studies such as [ 1]. 4.1 Performance Analysis of Cross-Lingual Relevance Classification Approaches Table2presents test performance metrics (accuracy and F1-score) for four cross- lingual relevance classification approaches, evaluated on a uniform test set of 750 tweets per language (English, Japanese, Korean, and Hindi) from original language data. The results, expressed with standard deviations across five runs, highlight the strengths and limitations of each approach, offering insights into their suitability for multilingual tweet analysis. Approach 1 (Language-Specific)leverages separate language-specific BERT models, trained on English-annotated tweets and their translations, with performance assessed on the test set. It achieves the highest accuracy and F1- score for English (97.72% ± 0.23% accuracy, 97.70% ± 0.23% F1-score), re- flecting the robustness of the BERT model on its native training language. For non-English languages, performance varies: Japanese (75.79% ± 0.49% accu- racy, 74.88% ± 0.56% F1-score), Korean (91.40% ± 0.31% accuracy, 91.35% ± 0.31% F1-score), and Hindi (78.08% ± 1.10% accuracy, 78.04% ± 1.10% F1- score) show that Korean benefits most from the language-specific model, while Towards Effective Topic Discovery in Multilingual Data9 Table 2: Test accuracy and F1-score for four cross-lingual relevance classification methods. Best results are shown in bold blue; second-best in bold. ApproachLanguage Model Accuracy (%) F1-score (%) Approach 1: Language-Specific ENBERT97.72 ± 0.2397.70 ± 0.23 JAJapanese BERT75.79 ± 0.49 74.88 ± 0.56 KOKorean BERT91.40 ± 0.3191.35 ± 0.31 HIHindi BERT78.08 ± 1.10 78.04 ± 1.10 Approach 2: English-only ENBERT97.72 ± 0.2397.70 ± 0.23 JABERT79.85 ± 0.6679.30 ± 0.69 KOBERT86.03 ± 0.22 85.78 ± 0.23 HIBERT90.59 ± 0.2790.53 ± 0.28 Approach 3: English Fine-Tuned mBERT ENmBERT96.68 ± 0.28 96.66 ± 0.28 JAmBERT50.21 ± 0.14 33.80 ± 0.30 KOmBERT50.56 ± 0.18 34.56 ± 0.39 HImBERT53.59 ± 0.28 40.97 ± 0.55 Approach 4: Hybrid ENmBERT94.81 ± 0.58 94.79 ± 0.58 JAmBERT78.49 ± 0.44 77.78 ± 0.49 KOmBERT83.57 ± 0.52 83.27 ± 0.54 HImBERT81.76 ± 0.48 81.18 ± 0.52 Japanese and Hindi exhibit lower and more variable performance, possibly due to translation errors or linguistic complexity. Approach 2 (English-only)relies on a single BERT model trained on English-annotated tweets, applied to the English portion of the test set and translated versions of non-English tweets. It excels in English (97.72% ± 0.23% accuracy, 97.70% ± 0.23% F1-score) and demonstrates strong performance on non-English languages: Japanese (79.85% ± 0.66% accuracy, 79.30% ± 0.69% F1-score), Korean (86.03% ± 0.22% accuracy, 85.78% ± 0.23% F1-score), and Hindi (90.59% ± 0.27% accuracy, 90.53% ± 0.28% F1-score). This approach outperforms Approach 1 for Japanese and Hindi, suggesting that translation to English provides a more consistent feature representation, likely due to high- quality translation and the effective model available in English language. Approach 3 (English Fine-Tuned mBERT)uses mBERT fine-tuned on English annotated tweets in a zero-shot setting. It performs well on English (96.68% ± 0.28% accuracy, 96.66% ± 0.28% F1-score) but shows significantly degraded performance on non-English languages: Japanese (50.21% ± 0.14% ac- curacy, 33.80% ± 0.30% F1-score), Korean (50.56% ± 0.18% accuracy, 34.56% ± 0.39% F1-score), and Hindi (53.59% ± 0.28% accuracy, 40.97% ± 0.55% F1- score). The low F1-scores (around 34-41%) compared to accuracy (50-54%) indi- cate a severe precision-recall imbalance, highlighting mBERTâs inability to gen- eralize from English to diverse linguistic structures without multilingual training. Approach 4 (Hybrid)enhances mBERT with a combined training set of English annotated tweets and their translations, improving cross-lingual perfor- mance. It achieves balanced results: English (94.81% ± 0.58% accuracy, 94.79% ± 0.58% F1-score), Japanese (78.49% ± 0.44% accuracy, 77.78% ± 0.49% F1- score), Korean (83.57% ± 0.52% accuracy, 83.27% ± 0.54% F1-score), and Hindi (81.76% ± 0.48% accuracy, 81.18% ± 0.52% F1-score). This approach out- 10Deepak Uniyal, Md Abul Bashar, and Richi Nayak Table 3: Distribution of Relevant (R) and Irrelevant (I) Tweets with Year-wise Percentage Breakdown (2013-2022) Across Four Classification Approaches App. LTotal R I â13 â14 â15 â16 â17 â18 â19 â20 â21 â22 #1 EN 4,769,850 47.80 52.20 3.33 4.01 5.47 5.24 5.96 6.32 4.77 14.88 24.12 25.91 JA 4,364,391 20.48 79.52 3.40 7.23 8.64 11.82 7.20 5.35 5.88 9.38 21.89 19.18 HI 27,258 35.44 64.56 0.50 1.22 1.01 2.49 2.52 5.16 3.21 7.75 26.44 49.70 KO 148,040 21.08 78.92 4.15 4.42 5.50 5.07 5.14 11.81 6.09 18.05 15.73 24.01 #2 EN 4,769,850 50.61 49.39 3.25 3.93 5.33 5.12 5.85 6.33 4.77 14.91 24.25 26.26 JA 2,459,340 16.16 83.84 3.29 6.55 8.95 10.00 6.38 5.01 7.35 10.01 22.32 20.05 HI 27,258 35.80 64.20 0.45 0.51 0.88 1.21 1.02 3.82 2.84 5.70 26.65 56.92 KO 148,040 17.92 82.08 4.37 4.86 6.11 5.62 5.86 12.25 5.92 17.32 15.50 22.18 #3 EN 4,769,850 48.76 51.24 3.30 3.99 5.44 5.22 5.98 6.28 4.69 14.88 24.20 26.02 JA 4,364,391 0.44 99.56 2.54 5.67 6.03 4.03 4.15 4.75 4.17 11.23 29.10 28.30 HI 27,258 5.01 94.99 0.29 0.07 1.03 0.51 0.66 1.98 1.03 4.18 21.10 69.15 KO 148,040 0.64 99.36 3.98 5.03 3.98 6.08 7.55 15.62 5.45 19.39 18.03 14.88 #4 EN 4,769,850 49.81 50.19 3.17 3.89 5.28 5.08 5.77 6.34 4.82 14.96 24.33 26.36 JA 4,364,391 21.50 78.50 3.36 7.16 8.63 8.35 6.20 4.85 10.51 10.25 21.84 18.82 HI 27,258 31.46 68.54 0.47 0.43 0.77 0.75 0.87 3.43 2.45 5.56 27.18 58.09 KO 148,040 20.13 79.87 4.92 4.43 5.46 5.38 5.58 12.05 6.17 17.48 15.63 22.89 Note:App.=Approach; L=Language; R=Relevant (%); I=Irrelevant (%); Years shown as â13ââ22 (2013â2022). All year columns show percentage of relevant tweets. performs Approach 3 across all non-English languages and shows competitive performance with Approach 1 for non-English languages, effectively bridging monolingual and multilingual strategies. In general, Approach 1 excels in English and Korean, but shows variability in Japanese and Hindi. Approach 2 provides the highest overall performance, particularly for English, Japanese, and Hindi, making it suitable for translation- based workflows. Approach 3 demonstrates the limitations of zero-shot multilin- gual transfer, while Approach 4 offers the balanced solution, leveraging multilin- gual training to improve consistency across languages. For practical deployment, Approach 2 is ideal for translation-heavy scenarios, whereas Approach 4 provides a robust foundation for truly multilingual applications, especially where native language data quality varies. 4.2 Temporal Distribution of Relevant Tweets Across Classification Approaches The temporal distribution of relevant tweets (Table3) reveals a marked increase in hydrogen energy discussions from 2013â22, with post-2019 acceleration across all languages aligning with global clean energy commitments like the EUâs Hy- drogen Strategy (2020) [ 24] and Japanâs National Hydrogen Strategy [25]. Language-specific patterns emerge clearly:English maintains the high- est relevance rates (47.80%â50.61%) with consistent growth peaking at 25.91%â Towards Effective Topic Discovery in Multilingual Data 11 Figure2: Topic Modelling 2013-2022 on a line graph showing the trending themes across multiple languages 26.36% in 2022. Japanese shows moderate relevance (12.36%â20.48%) with no- table 2021 peaks (21.89% in Approach 1). Korean exhibits similar patterns (17.92%â21.08% relevance) with delayed growth reaching 22.89%â24.01% by 2022. Hindi displays the most variable performance (5.01%â35.80% overall rele- vance) but dramatic 2022 surges (49.70%â69.15%), suggesting emerging market engagement despite Indiaâs complex multilingual social media landscape. Approach-specific insights:Approaches 1, 2, and 4 demonstrate bal- anced cross-lingual performance with steady temporal growth, while Approach 3 shows significant non-English underperformance (0.44% Japanese, 5.01% Hindi relevance), consistent with its lower classification accuracy noted in Table 2. The post-2020 acceleration, particularly the 2021-2022 peaks, aligns with ma- jor hydrogen policy initiatives, including Indiaâs 2020 National Green Hydrogen Mission [ 26]. These temporal patterns underscore hydrogen energyâs evolution from niche technology to mainstream policy priority, with language-specific adoption rates reflecting regional market maturity and policy engagement levels. 4.3 Topic Modeling The topic modelling analysis of hydrogen energy discourse from 2013 to 2022 across four languages (Figures2and3) reveals distinct thematic priorities and 12Deepak Uniyal, Md Abul Bashar, and Richi Nayak Figure3: Topic Modelling 2013-2022 on a bar graph showing the trending themes across multiple languages temporal evolution patterns under Approach 2âs translation-based classification framework. Dominant Topics and Language-Specific Patterns:English discourse is dominated by âhydrogen gasâ (T7, 629,378 tweets), representing 26.0% of to- tal discussions, followed by âhydrogen energy renewableâ (T4, 363,249 tweets, 15.1%) and âhydrogen productionâ (T10, 298,415 tweets, 12.4%). This distribu- tion reflects the global emphasis on hydrogen as a primary energy carrier and its scalability in production. Japanese discussions show a more balanced distribu- tion with âhydrogen energy renewableâ (T3, 224,413 tweets) leading at 27.3%, followed by âfirst hydrogen powerâ (T10, 141,273 tweets, 17.2%), indicating Japanâs focus on pioneering hydrogen technologies and renewable integration. Hindi exhibits significantly lower absolute volumes but concentrated thematic focus, with âgreen hydrogen Indiaâ (T3, 2,753 tweets, 28.2%) and âhydrogen fuel cellâ (T2, 1,860 tweets, 19.1%) dominating discussions, reflecting Indiaâs policy- driven approach to hydrogen adoption. âCar nitin gadkariâ (T1, 1,473 tweets, 15.09%) likely refers to the significant event when Indian Transport Minister Nitin Gadkari arrived at Parliament in a hydrogen-powered car in March 2022, generating substantial social media attention around hydrogen vehicle adoption in India [ 27]. Korean discourse emphasises âgreen hydrogen energyâ (T4, 6,091 tweets, 23%) and âgoing hydrogenâ (T10, 3,730 tweets, 14.1%), suggesting infrastructure development priorities. Temporal Evolution Patterns: Temporal trends reveal language-specific growth trajectories with notable post-2019 acceleration. English topics demon- strate consistent exponential growth, with âhydrogen gasâ (T7) and âgreen Towards Effective Topic Discovery in Multilingual Data 13 hydrogen projectâ (T2) showing dramatic increases from 2019-2022, reflecting global policy momentum. Japanese topics peaked in 2021, particularly âhydro- gen energy renewableâ (T3), likely coinciding with Tokyo Olympics hydrogen initiatives and national carbon neutrality commitments, before stabilising in 2022. Hindi shows the most dramatic late-stage growth, with several topics experiencing sharp increases in 2021-2022, particularly âgreen hydrogen Indiaâ (T3) jumping from minimal presence to over 1,000 tweets, aligning with Indiaâs National Hydrogen Mission launch. Korean trends display steady growth with âgreen hydrogen energyâ (T4) maintaining consistent upward trajectory through 2022, suggesting sustained industrial interest. Strategic Implications:The analysis reveals three distinct regional focus areas: global scalability (English emphasis on production and gas applications), technological leadership (Japanese focus on renewable integration and pioneering applications), and policy implementation (Hindi concentration on national mis- sions and Korean infrastructure development). These patterns suggest differenti- ated market opportunities, with English-speaking regions prioritising large-scale deployment, Japan leading in technology innovation, India focusing on policy- driven adoption, and Korea emphasising systematic infrastructure rollout. The temporal surge post-2020 across all languages, with peaks coinciding with major policy announcements (EU Hydrogen Strategy [24], Japanâs carbon neutrality pledge [28]), indicates coordinated global momentum in hydrogen energy transi- tion, offering strategic timing insights for market entry and investment decisions. 5 Discussion and Conclusion This study explored multilingual discussions on hydrogen energy by integrat- ing cross-lingual tweet classification with topic modeling, focusing on English, Japanese, Hindi, and Korean tweets from 2013 to 2022. We evaluated four cross- lingual relevance classification approaches, using language-specific, English-only, and multilingual models, and analysed relevant tweets to understand global clean energy conversations in multilingual digital spaces. Our results demonstrate that model performance varies significantly across approaches and languages, with the quality of translation playing a key role. Ap- proach 2 leveraging a single BERT model trained on English data and applied to translated tweets, achieved the highest overall performance for English and com- petitive results for Japanese, Korean, and Hindi. This suggests that high-quality translation to English provides a robust feature representation, outperforming Approach 1 for Japanese and Hindi, where native performance was lower (e.g., Japanese: 75.79% ± 0.49% accuracy, 74.88% ± 0.56% F1-score; Hindi: 78.08% ± 1.10% accuracy, 78.04% ± 1.10% F1-score). However, Approach 1 excelled in Korean (91.40% ± 0.31% accuracy, 91.35% ± 0.31% F1-score), indicating that language-specific models can be effective when trained and tested on high-quality native data. Topic modeling uncovered distinct regional discourse patterns such as En- glish emphasised production scalability (âhydrogen gasâ dominating at 26.2%), 14Deepak Uniyal, Md Abul Bashar, and Richi Nayak Japanese focused on renewable integration (âhydrogen energy renewableâ at 28.4%), Hindi concentrated on policy implementation, and Korean prioritised sustainable development. Temporal analysis showed coordinated global atten- tion on hydrogen energy after 2020, aligning with major policy initiatives, indi- cating market maturation, with classification accuracy directly influencing topic coherence quality. In future, we plan to explore sentiment trends over time, aspect-based sen- timent analysis for market insights, and improved cross-lingual methods that better handle native-language data instead of relying on only translations. Acknowledgements Funding The work has been partially supported by the Future Energy Exports CRC (w.fenex.org.au), whose activities are funded by the Australian Governmentâs Cooperative Research Centre Program. This is a FEnEx CRC Document 2025/ 22.RP4.0139.PHD-FNX-MILE0886. The authors would like to thank Queens- land University of Technology (QUT) for providing the High Performance Com- puting facilities used in this research and for their support. Publisher This preprint has not undergone peer review (when applicable) or any post- submission improvements or corrections. The Version of Record of this con- tribution is published in Data Science and Machine Learning. AusDM 2025. Communications in Computer and Information Science, vol 2765. Springer, Singapore., and is available online at https://doi.org/10.1007/978-981-95-6786- 7_34. References 1.D. Uniyal, R. Nayak, Twitterâs pulse on hydrogen energy in 280 characters: a data perspective, Social Network Analysis and Mining 14 (1) (2024) 37. 2.Y. Xu, H. Cao, W. Du, W. Wang, A survey of cross-lingual sentiment analysis: Methodologies, models and evaluations, Data Science and Engineering 7 (3) (2022) 279â299. 3.J. Devlin, M.-W. Chang, K. Lee, K. Toutanova, Bert: Pre-training of deep bidirec- tional transformers for language understanding, arXiv preprint arXiv:1810.04805 (2018). 4.S. Ruder, A. SĂžgaard, I. VuliÄ, Unsupervised cross-lingual representation learning, in: Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics: Tutorial Abstracts, 2019, p. 31â38. 5.C. Liu, W. Zhang, Y. Zhao, A. T. Luu, L. Bing, Is translation all you need? a study on solving multilingual tasks with large language models, arXiv preprint arXiv:2403.10258 (2024). Towards Effective Topic Discovery in Multilingual Data 15 6.T. Isbister, F. Carlsson, M. Sahlgren, Should we stop training more mono- lingual models, and simply use machine translation instead?, arXiv preprint arXiv:2104.10441 (2021). 7.P. Kumar, K. Pathania, B. Raman, Zero-shot learning based cross-lingual senti- ment analysis for sanskrit text with insuïŹicient labeled data, Applied Intelligence 53 (9) (2023) 10096â10113. 8.T. Pires, E. Schlinger, D. Garrette, How multilingual is multilingual bert?, arXiv preprint arXiv:1906.01502 (2019). 9.W. Yin, J. Hay, D. Roth, Benchmarking zero-shot text classification: Datasets, evaluation and entailment approach, arXiv preprint arXiv:1909.00161 (2019). 10.X. Wan, Co-training for cross-lingual sentiment classification, in: Proceedings of the Joint Conference of the 47th Annual Meeting of the ACL and the 4th Interna- tional Joint Conference on Natural Language Processing of the AFNLP, 2009, p. 235â243. 11.P. Prettenhofer, B. Stein, Cross-lingual adaptation using structural correspondence learning, ACM Transactions on Intelligent Systems and Technology (TIST) 3 (1) (2011) 1â22. 12.T. Mikolov, K. Chen, G. Corrado, J. Dean, EïŹicient estimation of word represen- tations in vector space, arXiv preprint arXiv:1301.3781 (2013). 13.A. Conneau, G. Lample, Cross-lingual language model pretraining, Advances in neural information processing systems 32 (2019). 14.A. Conneau, K. Khandelwal, N. Goyal, V. Chaudhary, G. Wenzek, F. GuzmĂĄn, E. Grave, M. Ott, L. Zettlemoyer, V. Stoyanov, Unsupervised cross-lingual repre- sentation learning at scale, arXiv preprint arXiv:1911.02116 (2019). 15.J. L. Fleiss, Measuring nominal scale agreement among many raters., Psychological bulletin 76 (5) (1971) 378. 16.M. R. Costa-JussĂ , J. Cross, O. Ăelebi, M. Elbayad, K. Heafield, K. Heffernan, E. Kalbassi, J. Lam, D. Licht, J. Maillard, et al., No language left behind: Scaling human-centered machine translation, arXiv preprint arXiv:2207.04672 (2022). 17.H. Zou, C. Caragea,JointMatch: A unified approach for diverse and collaborative pseudo-labeling to semi-supervised text classification, in: H. Bouamor, J. Pino, K. Bali (Eds.), Proceedings of the 2023 Conference on Empirical Methods in Nat- ural Language Processing, Association for Computational Linguistics, Singapore, 2023, p. 7290â7301.doi:10.18653/v1/2023.emnlp-main.451. URLhttps://aclanthology.org/2023.emnlp-main.451 18.T. NLP, Bert models for japanese text (clâtohoku/bertâjapanese),https:// github.com/cl-tohoku/bert-japanese , accessed: 2025-08-18 (2023). 19.J. Park, Distilkobert: Distillation of kobert,https://github.com/monologg/ DistilKoBERT(2019). 20.D. Kakwani, A. Kunchukuttan, S. Golla, G. N.C., A. Bhattacharyya, M. M. Khapra, P. Kumar, IndicNLPSuite: Monolingual Corpora, Evaluation Bench- marks and Pre-trained Multilingual Language Models for Indian Languages, in: Findings of EMNLP, 2020. 21.C. SĂĄnchez, H. Sarmiento, A. Abeliuk, J. PĂ©rez, B. Poblete, Cross-lingual and cross-domain crisis classification for low-resource scenarios, in: Proceedings of the International AAAI Conference on Web and Social Media, Vol. 17, 2023, p. 754â 765. 22.T. Balasubramaniam, R. Nayak, M. A. Bashar, Understanding the spatio-temporal topic dynamics of covid-19 using nonnegative tensor factorization: a case study, in: 2020 IEEE symposium series on computational intelligence (SSCI), IEEE, 2020, p. 1218â1225. 16Deepak Uniyal, Md Abul Bashar, and Richi Nayak 23.R. Egger, J. Yu, A topic modeling comparison between lda, nmf, top2vec, and bertopic to demystify twitter posts, Frontiers in sociology 7 (2022) 886498. 24.Euâs hydrogen strategy,https://energy.ec.europa.eu/topics/ energy-systems-integration/hydrogen_en, (Accessed: 2025-08-01) (2020). 25.Japan: First nation to form national hydrogen strategy,https://w. hydrogeninsight.com/, (Accessed: 2025-08-01) (2017). 26.National green hydrogen mission,https://mnre.gov.in/img/documents/ uploads/file_f-1673581748609.pdf, (Accessed: 2023-10-18) (2020). 27.Nitin gadkari using hydrogen-powered car,https://economictimes.indiatimes. com/, (Accessed: 2025-08-25) (2020). 28.Green growth strategy japan,https://w.meti.go.jp/english/policy/ energy_environment/global_warming/ggs2050/index.html, (Accessed: 2025-08- 25) (2020).