Multilingual sentiment analysis of summarized texts: a cross-language study of text shortening effects

Step-by-step description of the methodological pipeline (Fig. 1)

Figure 1 outlines a five-stage pipeline that was followed to ensure consistency and comparability across languages with different typological profiles.

Step 1: Data Collection. Six publicly available datasets were selected to ensure broad linguistic coverage and domain diversity. These datasets include IMDb (Maas et al., 2011), GermEval (Heinrich, Meyer & Gurevych, 2014), SemEval (Nakov et al., 2016), the Finnish Sentiment Dataset (Virtanen et al., 2019), the Hungarian Emotion Corpus (Singh, Gupta & Kumar, 2023), and ArSAS (Antoun, Baly & Hajj, 2020). To create a uniform binary classification setup, only texts with clearly labeled positive or negative sentiment were retained. The dataset distribution is presented in Table 1, showing that the final corpus spans eight languages with approximately 10,000 instances per language.

Step 2: Preprocessing. A standardized preprocessing pipeline was applied to all corpora to ensure uniformity across linguistic structures. The following operations were carried out:

• Text normalization involving case folding, Unicode standardization, and removal of special characters.

• Tokenization using SpaCy for European languages, and dedicated tokenizers for Finnish and Arabic to handle morphological and orthographic complexity.

• Stopword removal based on language-specific lists.

• Lemmatization for analytic and fusional languages, while morphological segmentation was prioritized for agglutinative languages such as Finnish and Hungarian.

• Diacritic normalization was carried out for Arabic to enhance lexical cohesion and reduce sparsity.

Step 3: Summarization. Three summarization paradigms were examined as part of the pipeline shown in Fig. 1:

• Extractive summarization was performed using TF-IDF based ranking and the TextRank algorithm (Mihalcea & Tarau, 2004) to identify salient content.

• Abstractive summarization was implemented using T5 (Raffel et al., 2020) and BART (Lewis et al., 2020), both fine-tuned on multilingual summarization datasets such as MLSum.

• A hybrid strategy was also introduced, whereby sentiment-bearing keyphrases were first extracted using TF-IDF combined with attention-based mechanisms, followed by controlled abstractive rephrasing to enhance fluency while preserving affective content (Singh, Gupta & Kumar, 2023; Scao et al., 2023).

Rationale for method selection

The choice of summarization techniques was guided by both methodological robustness and linguistic suitability. TF-IDF was selected for extractive summarization due to its ability to prioritize statistically distinctive, sentiment-bearing terms, a critical feature in languages where polarity is lexically encoded (e.g., English, German). TextRank (Mihalcea & Tarau, 2004) complements this by capturing discourse-level centrality through graph-based sentence similarity, preserving contextual coherence even in longer texts. Together, they offer a language-agnostic yet effective extractive baseline.

For abstractive summarization, T5 (Raffel et al., 2020) and BART (Lewis et al., 2020) were chosen as they represent the current state-of-the-art in sequence-to-sequence generation, support multilingual fine-tuning, and have demonstrated strong performance on the MLSum benchmark, a large-scale multilingual summarization corpus that aligns with our cross-linguistic design.

Our hybrid strategy builds on Singh, Gupta & Kumar (2023), who showed that combining extractive keyphrase selection with abstractive fluency improves sentiment retention. We enhanced this approach by integrating attention-based scoring from XLM-RoBERTa with TF-IDF to identify better sentiment-critical elements (e.g., negations, intensifiers, morphological markers), which are then protected during abstractive rewriting via soft lexical masking. This technique discourages the generator from altering flagged tokens.

For sentiment analysis, we employed mBERT and XLM-RoBERTa as multilingual baselines due to their broad cross-lingual transfer capabilities. Additionally, FinBERT (Virtanen et al., 2019) and AraBERT (Antoun, Baly & Hajj, 2020) were used for Finnish and Arabic, respectively, to explicitly model language-specific morphological patterns known to affect sentiment expression (e.g., Finnish case suffixes, Arabic root derivations).

Technical implementation and parameter selection: The summarization length was set to 30% of the original text, consistent with standard practices in the literature (Nenkova & McKeown, 2012) and validated in pilot studies as optimal for balancing compression and content retention. All hyperparameters were selected via grid search on language-specific validation sets (15% held-out per language) with early stopping (patience = 5 epochs). Specifically:

• TF-IDF and TextRank: Default configurations from Gensim and SpaCy were used as starting points, but term weighting schemes and the damping factor (for TextRank) were tuned to maximize ROUGE-1 on validation data.

• T5 and BART: Pretrained multilingual checkpoints (t5-base, facebook/bart-base) were fine-tuned with beam size = 4 (selected after testing {2, 4, 6} to maximize ROUGE without excessive repetition), maximum input length = 512 tokens, learning rate = $2\times {10}^{-5}$, and batch size = 16 values aligned with BERT-family fine-tuning conventions (Devlin et al., 2019) and GPU memory constraints (NVIDIA V100 32 GB).

Full hyperparameter grids and validation curves are available in our public repository.

Step 4: Sentiment Analysis. Sentiment classification was conducted using both multilingual and language-specific transformer-based models:

• Multilingual models: mBERT (Devlin et al., 2019) and XLM-RoBERTa (Conneau et al., 2020) were employed as baselines for cross-lingual transfer.

• Language-specific models: FinBERT (Virtanen et al., 2019) for Finnish and AraBERT (Antoun, Baly & Hajj, 2020) for Arabic were used to capture morphological and lexical nuances.

All models were fine-tuned for binary classification using the AdamW optimizer with a learning rate of $2\times {10}^{-5}$ and batch size of 16. Early stopping was employed to prevent overfitting, using validation loss as the stopping criterion.

Step 5: Evaluation. The impact of summarization on sentiment classification was evaluated using three distinct metric categories:

Sentiment classification metrics: Accuracy, Precision, Recall, F1-score, and sentiment consistency (defined as the percentage of samples where the sentiment prediction remains unchanged after summarization).

Summarization quality metrics: ROUGE-1, ROUGE-2, ROUGE-L (Lin, 2004), and BLEU (Papineni et al., 2002), which evaluate informativeness and fluency.

Semantic drift detection: Cosine similarity was computed between sentence embeddings from XLM-RoBERTa to quantify changes in semantic content (Zhang, Li & Wang, 2023).

To ensure the validity of findings, all experiments were repeated five times using different random seeds with a standardized 70%/15%/15% (train/validation/test) split ratio. Stratified sampling was applied to maintain class balance across all subsets. Where appropriate, 5-fold cross-validation was performed during hyperparameter tuning, and paired t-tests ( $\alpha =0.05$) were used to evaluate statistical significance.

This comprehensive methodology facilitates reproducible and cross-linguistically comparable evaluation of how summarization techniques affect sentiment analysis. Its structured design supports language-agnostic conclusions while remaining sensitive to morphological and syntactic variation.

Overview of the obtained results

The impact of text summarization on sentiment classification accuracy was systematically evaluated across eight typologically diverse languages, with results indicating a significant dependence on both the summarization method employed and the morphological complexity of the language. A consistent decline in sentiment classification performance was observed across all languages following summarization, with the magnitude of degradation varying according to linguistic structure.

In languages with simpler morphological systems, such as English, German, and Spanish, modest reductions in accuracy were recorded, ranging from 2.6 to 3.8 percentage points under extractive methods. In contrast, morphologically rich languages, including Finnish, Hungarian, and Arabic, exhibited substantially greater declines, particularly when abstractive summarization was applied. For instance, accuracy drops of 6.9 and 6.1 percentage points were observed in Finnish and Arabic, respectively, under abstractive approaches, compared to baseline performance.

Extractive summarization was found to preserve sentiment more effectively than abstractive methods, as evidenced by higher scores in accuracy, F1-score, and sentiment consistency. This preservation is attributed to the retention of original lexical and syntactic structures, which are critical for maintaining affective cues in morphologically complex systems. Abstractive summarization, while achieving competitive fluency as measured by ROUGE and BLEU metrics, introduced measurable semantic drift, leading to polarity shifts and reduced classification reliability, especially in agglutinative and inflectional languages.

A hybrid approach, combining extractive selection of sentiment-bearing phrases with controlled abstractive rewriting, was shown to mitigate sentiment distortion while maintaining textual coherence. In Arabic, this method reduced accuracy loss by 4.2 percentage points compared to pure abstractive summarization, demonstrating its potential for balancing fluency and emotional fidelity.

Evaluation metrics, including ROUGE-1, BLEU, precision, recall, and F1-score, were consistently applied across all languages and methods, enabling a standardized, cross-linguistic comparison. Sentiment consistency, defined as the proportion of samples with unchanged predicted polarity post-summarization, was introduced as a key indicator of affective stability, revealing a strong correlation between structural preservation and sentiment retention.

These findings confirm that the interaction between summarization strategy and linguistic typology plays a critical role in determining sentiment analysis outcomes. The results further highlight the necessity of methodological alignment with language-specific characteristics, particularly in multilingual NLP applications where accurate sentiment transmission is essential.

Sentiment analysis before summarization

The sentiment classification results before summarization indicated high accuracy levels across all languages. A baseline evaluation was conducted using multilingual models such as mBERT and XLM-RoBERTa, alongside language-specific models like FinBERT and AraBERT for Finnish and Arabic, respectively. The classification performance metrics for each language are presented in Table 2.

Before text summarization, sentiment classification metrics were obtained for multiple languages. The highest accuracy was achieved for English (92.15%), followed by French (89.55%) and German (88.25%). In Spanish and Italian, accuracy values were 87.05% and 86.35%, respectively. For Finnish, Hungarian, and Arabic, accuracy was lower, ranging from 84.35% to 81.75%. A similar trend was observed across Precision, Recall, and F1-score metrics, with English performing best and Arabic lowest. Overall, Latin and Germanic languages demonstrated stronger baseline results, while Finno-Ugric and Semitic languages trailed behind.

Summarization results

The effects of summarization on sentiment classification were analyzed separately for extractive, abstractive, and hybrid methods. Summarization quality was assessed using ROUGE-1 and BLEU metrics, which measure content overlap and fluency, respectively.

Extractive summarization results

Extractive summarization, which selects key sentences from the original text, demonstrated superior sentiment preservation, particularly in morphologically rich languages. The ROUGE-1 and BLEU scores for extractive summarization are presented in Table 3.

As shown in Table 3, English and French achieved the highest ROUGE-1 scores (89.3% and 88.5%), indicating strong content retention. In contrast, Finnish and Arabic scored the lowest (78.9% and 75.3%), reflecting greater information loss in morphologically complex languages. BLEU results followed a similar trend, confirming better structural preservation in languages with simpler grammatical structures.

Abstractive summarization results

Abstractive summarization, which generates new phrasing through paraphrasing, introduced greater variability and semantic drift, especially in agglutinative and inflectional languages. The respective metrics are provided in Table 4.

As shown in Table 4, abstractive summarization yielded lower ROUGE-1 and BLEU scores compared to extractive methods across all languages. While English and French maintained relatively high scores (ROUGE-1: 87.1% and 86.3%; BLEU: 83.2% and 82.1%), the decline was more pronounced in morphologically complex languages. For example, Arabic exhibited the lowest ROUGE-1 (70.1%) and BLEU (66.3%) scores, indicating significant challenges in preserving both content and fluency during paraphrasing.

Hybrid summarization results

To address the trade-off between sentiment retention and textual coherence, a hybrid summarization method was developed, combining extractive selection of sentiment-bearing keyphrases with controlled abstractive rewriting. The approach first identifies emotionally salient phrases using TF-IDF weighting and attention-based scoring, followed by fine-tuned T5 and BART models to generate fluent paraphrases. This design aims to preserve critical morphological and syntactic markers such as negations, intensifiers, and inflectional suffixes, while enhancing readability.

As shown in Table 5, the hybrid method achieved consistently high performance across all languages. In English and French, ROUGE-1 scores exceeded 86%, while in morphologically complex languages like Arabic and Finnish, the hybrid approach outperformed both extractive and abstractive baselines in fluency and content overlap.

Crucially, the hybrid method significantly improved sentiment preservation. In Arabic, while abstractive summarization reduced sentiment accuracy from 81.75% to 75.65%, the hybrid model maintained an accuracy of 78.85%, representing a +4.2 percentage point improvement. Similarly, in Finnish, the hybrid approach reduced the accuracy loss from 6.9% (abstractive) to 4.1%, effectively preserving affective suffixes such as -han and -pa.

To quantify semantic drift (defined as preserved factual content with shifted emotional tone (Li et al., 2021)), XLM-RoBERTa embeddings analysis confirmed a statistically significant 15% lower cosine distance shift in hybrid approaches ( $p<0.01$). Concrete examples included:

• Finnish: Omission of the emphatic suffix -han (e.g., “Hyvseon” $\to$ “Hyv on”) reduced positive emphasis (sentiment score decrease from 0.78 to 0.52).

• Hungarian: Word order inversion (“Nem szeretem” [I don’t like] $\to$ “Szeretem nem” [I like not]) flipped sentiment (accuracy drop 12%).

Qualitative Analysis of Best and Worst Cases. To further understand the mechanisms of sentiment preservation and distortion, we conducted a detailed qualitative analysis of the best-performing and worst-performing samples across languages. Cases were selected based on prediction confidence and label mismatch between original and summarized texts, then annotated by two native speakers for linguistic causes.

• Best case (German—Hybrid method): Original: “Der Film war wirklich hervorragend, obwohl die Handlung nicht perfekt war!” (“The movie was truly excellent, although the plot was not perfect!”) Hybrid summary: “Film war wirklich hervorragend” $\to$ Sentiment preserved (positive). Why it succeeded: The hybrid method retained the intensifier “wirklich” and positive adjective “hervorragend”, while correctly omitting the concessive clause without altering polarity.

• Worst case (Finnish—Abstractive method): Original: “Elokuva olihan hyvä, mutta tylsä.” (“The movie was, after all, good, but boring.” – nuanced negative) Abstractive summary: “Elokuva oli hyvä.” (“The movie was good.”) $\to$ Misclassified as positive. Why it failed: The model dropped the concessive particle “-han” (implying doubt) and the contrastive “mutta”, flattening polarity. Extractive and hybrid methods retained both.

• Structural error (Hungarian): Abstractive reordering “Nem szeretem ezt” $\to$ “Ezt szeretem nem” inverted emphasis and flipped sentiment in 12% of similar cases (validated via native annotation). Extractive methods preserved original word order and polarity.

These cases confirm that morphological fidelity (preserving particles, suffixes) and syntactic integrity (maintaining clause structure and word order) are decisive for sentiment accuracy—more so than lexical coverage alone.

Figure 2 illustrates the comparative performance across methods. Extractive summarization (blue line) achieved the highest ROUGE-1 and BLEU scores, followed closely by the hybrid method (green), while abstractive summarization (orange) consistently underperformed. The performance gap widened in morphologically complex languages, where hybridization bridged approximately 50% of the fluency-accuracy trade-off, demonstrating its effectiveness in balancing content fidelity and linguistic coherence.

Sentiment analysis after summarization

The impact of summarization on sentiment classification accuracy was evaluated across all three methods. Table 6 presents the post-summarization sentiment metrics.

A consistent decline in sentiment accuracy was observed across all languages, with the most significant drops occurring in morphologically rich languages. Extractive summarization preserved sentiment most effectively, as evidenced by higher F1-scores and accuracy values. Abstractive methods introduced notable distortion, particularly in Arabic, where accuracy decreased by 6.1% compared to the baseline, and in Hungarian, with a 7.5% drop. Although the hybrid method is not aggregated in this table, it demonstrated superior performance in targeted evaluations: it reduced polarity shifts and outperformed the abstractive baseline by 4.2 percentage points in F1-score for Arabic.

These findings confirm that summarization method and morphological complexity are key determinants of sentiment preservation. The hybrid approach emerges as a robust solution for multilingual sentiment-sensitive applications, offering a balanced trade-off between fluency, content retention, and emotional fidelity.

Comparison of sentiment metrics before and after summarization

A decline in sentiment classification performance was observed across all languages following text summarization, with the extent of degradation influenced by both the morphological complexity of the language and the specific summarization method applied. Extractive, abstractive, and hybrid approaches were evaluated using precision, recall, F1-score, and accuracy, with results systematically compared to baseline performance prior to summarization.

As illustrated in Fig. 3, sentiment accuracy decreased across all languages, with the most pronounced drops occurring in morphologically complex systems such as Finnish, Hungarian, and Arabic. Extractive summarization demonstrated superior preservation, with accuracy in English decreasing only from 92.15% to 89.55%, and in German from 88.25% to 85.65%. In contrast, abstractive methods led to greater losses: Arabic experienced a drop from 81.75% to 75.65%, while Finnish and Hungarian exhibited reductions of 8.5 and 7.5 percentage points, respectively. The hybrid approach mitigated these losses, maintaining accuracy at 78.85% in Arabic and 80.25% in Finnish, representing a significant improvement over pure abstractive summarization.

Precision, as shown in Fig. 4, followed a similar trend. The highest pre-summarization precision was observed in English (90.45%), while Arabic recorded the lowest (77.65%). After summarization, extractive methods preserved precision most effectively, with only minor reductions (e.g., 90.45% to 87.85% in English). Abstractive summarization resulted in more substantial declines, particularly in Finnish and Arabic, where key sentiment-bearing morphemes were omitted during paraphrasing. The hybrid method reduced precision loss by approximately 3.5% compared to abstractive outputs in Arabic.

Recall, depicted in Fig. 5, demonstrated consistent degradation across all methods, with abstractive summarization exhibiting the steepest declines in Finnish and Hungarian. This indicates a failure to retrieve sentiment-relevant content during generation, likely due to structural reformulation and semantic drift. Extractive methods retained higher recall by preserving original sentence segments, while the hybrid strategy improved recall by 4.2% in Arabic compared to abstractive summarization.

The F1-score, which balances precision and recall, is presented in Fig. 6. Extractive summarization achieved the highest F1-scores post-summarization, with English maintaining 88.25% and Arabic 71.85%. Abstractive methods showed the lowest F1 performance, especially in Arabic (70.95%) and Finnish (73.45%), confirming the trade-off between fluency and emotional fidelity. Notably, the hybrid method achieved a 4.2% higher F1-score in Arabic compared to abstractive summarization, demonstrating its effectiveness in preserving sentiment while enhancing readability.

Collectively, the results confirm that extractive summarization is the most effective in preserving sentiment, particularly in languages where inflectional morphology and syntactic structure contribute to affective meaning. Abstractive methods, while generating more fluent summaries, introduce semantic drift that disproportionately affects morphologically rich languages. The hybrid approach emerges as a robust alternative, combining the structural fidelity of extractive selection with the fluency of abstractive rewriting, thereby reducing sentiment distortion and improving classification stability. These findings underscore the necessity of method selection based on linguistic typology, especially in multilingual applications such as social media monitoring and cross-lingual opinion mining.

Summary of findings

The results demonstrate that text summarization has a measurable impact on sentiment classification accuracy, with the degree of distortion strongly influenced by the morphological and syntactic characteristics of the language. Extractive summarization consistently outperforms abstractive methods in preserving sentiment, particularly in morphologically rich languages such as Finnish, Hungarian, and Arabic. In these languages, the retention of original sentence structures, inflectional markers, and syntactic configurations enables higher sentiment consistency, with accuracy drops of 6.9–10.3 percentage points observed under abstractive summarization compared to 2.9–6.8 points under extractive approaches.

In contrast, languages with simpler morphological systems, such as English, German, and French, experience more modest declines in sentiment accuracy post-summarization. For English, extractive summarization resulted in a 2.6-percentage-point drop in accuracy (from 92.15% to 89.55%), while abstractive summarization led to a 6.5-point decrease. These findings suggest that linguistic structure plays a critical role in determining the extent of sentiment distortion during text shortening.

Abstractive summarization improves textual fluency, as reflected in ROUGE and BLEU scores, but introduces semantic drift that disproportionately affects sentiment classification in morphologically complex languages. This is particularly evident in Finnish and Arabic, where the omission or alteration of case markers, suffixes, and root-based inflections during paraphrasing leads to polarity shifts and reduced classification performance.

A hybrid summarization approach combining extractive selection of sentiment-bearing keyphrases (via TF-IDF and attention mechanisms) with controlled abstractive rewriting demonstrates a favorable balance between fluency and affective fidelity. In Arabic, this method achieved an accuracy of 78.85%, representing a +3.2-percentage-point improvement over pure abstractive summarization and reducing the baseline loss by 50%. Similarly, in Finnish, the hybrid model limited accuracy degradation to 4.1 percentage points, significantly outperforming abstractive methods. Semantic drift analysis based on XLM-RoBERTa embeddings confirmed a 15% lower average cosine distance shift in hybrid outputs, indicating enhanced stability in semantic and emotional content.

These findings underscore the importance of aligning summarization strategies with linguistic typology. Extractive methods are recommended for high-stakes sentiment-sensitive applications in morphologically complex languages, while hybrid approaches offer a robust alternative for scenarios requiring both readability and sentiment preservation.

Conclusions

The influence of text summarization on sentiment analysis accuracy was found to be significantly modulated by linguistic typology, particularly morphological complexity and syntactic flexibility. Across eight typologically diverse languages, a consistent decline in sentiment classification performance was observed following summarization, with the magnitude of degradation varying systematically with language structure.

In morphologically simple languages such as English, German, and Spanish, only minor reductions in accuracy were recorded, ranging from 2.6% to 4.5%. These results suggest that analytic and fusional languages are relatively resilient to affective distortion during text shortening, as key sentiment cues are often lexically encoded and less dependent on inflectional morphology.

In contrast, morphologically rich languages, including Finnish, Hungarian, and Arabic, exhibited substantially greater sentiment distortion, with accuracy losses between 8% and 12%. These effects were particularly pronounced under abstractive summarization, where paraphrasing led to the omission or alteration of case markers, suffixes, diacritics, and root-based derivations that encode emotional valence. The vulnerability of agglutinative and inflectional systems to structural reformulation was confirmed through both classification metrics and semantic drift analysis.

Extractive summarization was shown to preserve sentiment more effectively than abstractive methods, particularly in languages where syntactic and morphological integrity is critical for affective expression. This preservation is attributed to the retention of original lexical units and grammatical structures, minimizing the risk of polarity shifts. Abstractive approaches, while achieving higher fluency as measured by ROUGE and BLEU, introduced semantic drift that disproportionately impacted sentiment accuracy in complex linguistic systems.

Language-specific models, FinBERT for Finnish and AraBERT for Arabic, demonstrated superior performance compared to multilingual baselines (mBERT, XLM-RoBERTa), with gains of 4–7 percentage points in F1-score. These results highlight the advantages of domain-adapted pretraining and morphological awareness in sentiment-sensitive tasks.

A hybrid summarization strategy, combining extractive selection of sentiment-bearing keyphrases (via TF-IDF and attention mechanisms) with controlled abstractive rewriting, was found to mitigate sentiment distortion while maintaining textual coherence. In Arabic, this approach reduced accuracy loss by 4.2 percentage points relative to pure abstractive summarization, and in Finnish, it limited degradation to 4.1 points, outperforming both standalone methods. Semantic stability, measured via cosine similarity of XLM-RoBERTa embeddings, was 15% higher in hybrid outputs, indicating improved affective fidelity.

These findings confirm that both the choice of summarization method and the linguistic properties of the target language play decisive roles in sentiment preservation. High data quality, model specificity, and structural alignment are essential for minimizing affective distortion in multilingual NLP pipelines.

This study presents the first large-scale, systematic cross-linguistic evaluation of sentiment preservation under text summarization across eight typologically diverse languages. By integrating morphological typology, language-specific models, and hybrid summarization strategies, we establish a new benchmark for affect-aware multilingual text compression, addressing a significant gap in current NLP research.

Recommendations

Based on the findings, the following recommendations are proposed to enhance sentiment preservation in multilingual summarization systems:

• 1.

  Adoption of Hybrid Summarization Strategies: A combination of extractive selection of sentiment critical phrases and controlled abstractive rewriting is recommended, particularly for morphologically complex languages. This approach balances fluency with emotional fidelity and has been shown to reduce sentiment degradation by up to 50% compared to abstractive methods. This improvement exceeds the 3.1% F1-score gain reported by Singh, Gupta & Kumar (2023) for a similar hybrid strategy in low-resource sentiment datasets, demonstrating the added value of attention-guided keyphrase masking in morphologically complex contexts.

• 2.

  Use of Language-Specific Models: For languages with rich morphology or low-resource status, the deployment of specialized models such as FinBERT or AraBERT is advised. Where such models are unavailable, fine-tuning multilingual transformers (e.g., XLM-RoBERTa) on domain-specific, sentiment-annotated corpora is recommended to improve affective sensitivity.

• 3.

  Integration of Sentiment-Aware Training Objectives: Summarization models should be fine-tuned using sentiment-labeled data, with explicit optimization for polarity consistency. Attention mechanisms targeting sentiment-bearing morphemes (e.g., negations, intensifiers, case markers) can be incorporated to preserve emotional cues during compression.

• 4.

  Development of Sentiment-Sensitive Evaluation Metrics: Current metrics, such as ROUGE and BLEU, are insufficient for assessing affective fidelity. The development and adoption of sentiment-aware evaluation protocols incorporating polarity consistency, emotion alignment, and cultural context are strongly encouraged to enable more reliable assessment of summarized outputs.

• 5.

  Incorporation of Cultural and Contextual Factors: Sentiment expression varies across sociocultural contexts. Model training and evaluation should include culturally adapted datasets to reflect language-specific norms of emotional expression, particularly in applications such as social media monitoring and customer feedback analysis.

• 6.

  Exploration of Next-Generation LLMs with Controlled Generation: Large language models (e.g., BLOOM, LLaMA 2) offer multilingual capabilities but require further investigation into their ability to preserve sentiment. Techniques for controlling output for emotional consistency, mitigating hallucination, and reducing cultural bias should be prioritized in future work.

• 7.

  Implementation in Real-World Multilingual Applications: The proposed hybrid and language-specific strategies are recommended for deployment in high-stakes domains such as cross-lingual opinion mining, market analysis, and public sentiment monitoring, where accurate transmission of affective content is critical.

These recommendations are intended to guide the design of more robust, equitable, and sentiment-aware NLP systems, particularly in multilingual environments where linguistic diversity poses significant challenges to automated affect analysis.