This paper presents a competitive approach to multilingual subjectivity detection using large language models (LLMs) with few-shot prompting. We participated in Task 1: Subjectivity of the CheckThat! 2025 evaluation campaign. We show that LLMs, when paired with carefully designed prompts, can match or outperform fine-tuned smaller language models (SLMs), particularly in noisy or low-quality data settings. Despite experimenting with advanced prompt engineering techniques-such as debating LLMs and various example selection strategies-we found limited benefit beyond well-crafted standard few-shot prompts. Our system achieved top rankings across multiple languages in the CheckThat! 2025 subjectivity detection task, including first place in Arabic and Polish, and top-four finishes in Italian, English, German, and multilingual tracks. Notably, our method proved especially robust on the Arabic dataset, likely due to its resilience to annotation inconsistencies. These findings highlight the efectiveness and adaptability of LLM-based few-shot learning for multilingual sentiment tasks, ofering a strong alternative to traditional fine-tuning, particularly when labeled data is scarce or inconsistent.
CEUR
ceur-ws.org work aims to demonstrate that LLMs can indeed compete with, and potentially outperform, fine-tuned SLMs in subjectivity detection tasks through the application of techniques such as careful prompting and few-shot prompting. We investigate the extent to which precise prompt engineering and the provision of relevant examples within the prompt can enhance the performance of LLMs, thereby showcasing their potential for robust and adaptable subjectivity detection in academic and real-world settings. While these techniques show promise in theory, our experiments reveal that they do not consistently improve performance on the tested datasets, suggesting important limitations and directions for further research.
The computational detection of subjective language has evolved significantly from its early rule-based foundations to contemporary data-driven approaches. This progression reflects broader trends in natural language processing while addressing challenges in identifying opinionated content. Current research emphasizes two critical aspects: (1) the development of increasingly sophisticated models capable of capturing linguistic nuance [9, 10], and (2) the creation of specialized techniques to optimize these models for subjectivity analysis [11, 12, 13].
Architectural Evolution. Traditional lexicon-based and supervised learning approaches have given
way to transformer-based models, with BERT-style architectures demonstrating strong performance on
binary subjectivity classification [ 14, 15, 16]. However, the emergence of LLMs has introduced new
capabilities in detecting implicit subjectivity through contextual reasoning. Indeed, a key advantage of
LLMs is their advanced ability to recognize subtle linguistic cues like irony, sarcasm, or implicit bias [17].
Their architectural design enables them to capture complex dependencies within sequential data, leading
to a deeper understanding of intricate relationships between words and emotions. Recent empirical
studies have demonstrated that LLMs consistently achieve higher overall accuracy in sentiment analysis,
often outperforming specialized pre-trained transformer models due to their comprehensive grasp of
human thought and emotion [10]. This ability indicates strong potential for subjectivity detection, as
recognizing nuanced evaluative language is key to distinguishing subjective from objective statements
[
Prompting Strategies. Despite their inherent capabilities, optimizing LLM performance in specialized tasks like subjectivity detection often requires appropriate prompting. While a wide range of prompting strategies exists, we focus in this work on a selected subset of strategies which we evaluate in our experiments.
• Prompt engineering involves meticulously designing input queries to guide LLMs toward
desired outputs. This technique is crucial for clearly defining the nuances of subjectivity for the
model, specifying output formats, and aligning the LLM’s reasoning with task-specific objectives
[18, 19]. However, prompt efectiveness can be highly sensitive to wording, and small changes
may lead to inconsistent results [20, 21].
• In-context Learning (ICL) refers to the ability of LLMs to perform tasks by conditioning on
input-output examples provided directly in the prompt, without updating model parameters. A
common subcategory of ICL is few-shot learning, where the prompt includes a small number
of labeled examples to help the model infer the task and classification criteria [
The CheckThat! lab [
We utilize the provided multilingual dataset, which comprises sentence-level annotations labeled as either OBJ (objective) or SUBJ (subjective). Table 1 summarizes the number of annotated sentences per language and split. The dataset exhibits class imbalance across languages and splits, with some languages (e.g., Italian and Arabic) showing a predominance of OBJ labels, while others (e.g., Bulgarian) present a more balanced distribution.
During the exploratory phase, we focus primarily on English and Arabic. These two languages were selected due to their difering class distributions and dataset sizes. We operate under the assumption that insights obtained from these languages are transferable to the other languages in the dataset.
We explore several strategies to improve subjective sentence classification. We experiment with the following three main approaches: prompt engineering, few-shot learning, and multi-agent LLM setups. We also use fine-tuned SLMs as baselines.
We systematically evaluate the impact of prompt phrasing and label framing on classification performance. We compare: • A minimal, generic prompt vs. a detailed one generated from the annotation guidelines. • Label framing using explicit terms (“Subjective”/“Objective”) vs. neutral terms (“Category 0”/“Category 1”).
• Binary yes/no questions (e.g., “Is the sentence subjective?”) as an alternative to direct classification. These variations are designed to probe how linguistic framing afects the model’s interpretability and consistency.
We experiment with a range of few-shot prompting configurations to assess how the number and selection of support examples influence performance. Specifically, we compare: • Prompting setups with 0 (no-shot), 6-shot, and 12-shot examples. • Example selection strategies based on: (a) Semantic similarity, (b) Semantic dissimilarity, and (c)
Random sampling.
For the semantic approaches, similarity and dissimilarity are measured using cosine similarity between sentence embeddings generated by OpenAI’s text-embedding-3-small model. To ensure class balance in the few-shot prompts, we selected an equal number of examples from each class (subjective and objective). For instance, in the 6-shot setting, the prompt included the three most similar (or dissimilar) subjective examples and the three most similar (or dissimilar) objective ones. This balance helps prevent prompt-induced bias toward a particular class label.
We design a set of multi-agent prompting experiments to investigate the interpretability and robustness of LLM outputs: • Debate setup: Two agents argue why a sentence is subjective vs. objective; a third model acts as a judge. • Adversarial reasoning: One agent argues why the sentence is not subjective and another why it is not objective, and a judge makes the final call. • Extended framing: We include all four perspectives—Subjective, Not Subjective, Objective, and
Not Objective— and a judge making the final decision.
This section presents a detailed evaluation of multiple modeling strategies for subjectivity classification, including fine-tuned transformers, prompted LLMs with and without few-shot examples, advanced prompt reframing, and agent-based debating approaches. Given the dataset’s imbalance, the oficial primary evaluation metric is the macro-averaged F1-score, which equally weights both classes. We also pay particular attention to SUBJ recall, as the subjective class is often underrepresented and may be more informative in downstream analyses. This imbalance was mitigated using a weighted BCE loss, warmup training, and early stopping to ensure stability and generalizability. We first report results with diferent system variants on the development subset and then discuss the oficial results obtained with the 2025 test set.
Table 2 summarizes the performance of supervised transformer models trained on English and Arabic data. RoBERTa-Base, fine-tuned on English data, achieves the best performance overall, with a macro F1 of 0.70 and a notably high macro precision of 0.79. However, the model struggles with subjective instances, achieving only 0.39 recall for the subjective class, which suggests a strong bias toward the majority (objective) class.
In Arabic, the results are considerably weaker across the board. While BERT-Base-Arabertv02 achieves the best Arabic macro F1 score (0.55), subjective recall remains modest (0.47). Despite the use of language-specific models and XLM-RoBERTa for cross-lingual encoding, the performance gap between English and Arabic remains substantial.
We further compare several strategies for selecting few-shot examples: random sampling, similaritybased sampling, and dissimilarity-based sampling. In similarity-based sampling, we choose examples that are most similar to the test sentence, while in dissimilarity-based sampling, we select those that are the most diferent. Similarity is measured using the cosine similarity between sentence embeddings generated by GPT-3. Results are shown in Table 4.
Interestingly, random selection outperforms similarity-based strategies across all models. For
GPT4o-mini, random sampling yields the best macro F1 (0.76), whereas dissimilarity-based selection ofers
a better recall (0.73) but slightly lower overall F1. A similar trend is observed for Qwen-72B, where
dissimilar sampling boosts recall (+0.07 over similarity) but ofers minimal F1 gain. This suggests that
dissimilar examples may help capture broader linguistic variance, aiding generalization. These results
contrast with earlier findings highlighting the benefits of semantically similar examplars for in-context
learning [
We investigate whether the way labels are framed afects model behavior. Reframing “subjective vs. objective” as a binary question (e.g., “Is the sentence subjective? Yes/No”) or as category labels (“Category 1 vs. Category 2”) leads to slight F1 gains over the base prompt (Table 5). Framing clearly influences the model’s inductive bias, with category labels yielding better subjective precision (0.69) and macro F1 (0.72).
Debating-based prompting (Table 6) also provides strong results. The setup where one LLM argues for subjectivity, another for objectivity, and a judge decides, achieves the best macro F1 overall (0.77). Notably, this format significantly enhances subjective recall (up to 0.74), suggesting that reasoningfocused prompting facilitates more balanced decisions. Debate variants using negated prompts (e.g., “Not Subjective” vs. “Not Objective”) also perform competitively.
Finally, we evaluate an ensemble voting strategy that aggregates predictions from five diverse models: RoBERTa-Base, GPT-4o-mini, LLaMA 70B, Qwen 72B, and Aya-Expanse 32B. As shown in Table 7, this ensemble achieves the highest overall macro F1 score (0.79), with a strong subjective precision of 0.77. These results indicate that ensembling models with heterogeneous architectures and training paradigms can efectively capture complementary perspectives on subjectivity, enhancing robustness and performance.
The oficial evaluation of the CheckThat! 2025 campaign comprised three settings: • Monolingual: train and test on data in a given language (Arabic, Italian, German, English). • Multilingual: train and test on data comprising several languages. • Zero-shot: train on several languages and test on unseen languages (Romanian, Polish, Ukrainian,
Greek).
For our final submitted system in the oficial campaign evaluation, we adopted the extended-prompt strategy using randomly selected 6-shot examples, paired with an ensemble of multiple models, including GPT-4 variants (GPT-4o-mini, GPT-4.1-mini), RoBERTa, LLaMA 70B, and Qwen 72B. In the zero-shot setting, the in-context examples were provided in English, following the task guidelines.
Final results are reported in Table 8.
Our team demonstrated strong results across multiple languages, achieving first place in Arabic and Polish, and top-three positions in the majority of the evaluated languages, including Italian, English, and multilingual settings. This consistent performance underscores the robustness and generalizability of our approach using LLM-based few-shot learning.
Our experiments demonstrate that leveraging large language models (LLMs) instead of fine-tuned smaller language models (SLMs) can yield highly competitive results across multiple languages and settings. However, the efectiveness of LLMs critically depends on the quality of prompt design. For instance, advanced prompting strategies such as debating LLMs, where multiple model outputs are cross-examined, did not lead to substantial improvements over standard few-shot prompting. Similarly, varying the example selection method, whether by similarity, dissimilarity, or random choice, showed no significant impact on final performance. These findings suggest that while prompt engineering remains essential, more complex example selection or ensemble strategies may not always provide additional gains.
The most notable result was in Arabic, where we outperformed the second-ranked team by a
substantial margin of +0.10 Macro F1-score. We attribute this advantage partly to the nature of the
Arabic dataset, which exhibits annotation inconsistencies. Unlike fine-tuned models that heavily depend
on high-quality labeled training data, our few-shot LLM approach is less afected by such noise. Prior
research has indicated that in-context learning with LLMs can be relatively independent of the exact
label quality provided in training examples [
Overall, our findings suggest that LLMs, combined with carefully crafted few-shot prompts, ofer a powerful and flexible alternative to traditional fine-tuning approaches, especially when training data quality varies. This has important implications for future multilingual sentiment analysis tasks and other NLP challenges where data quality and multilingual coverage are key concerns.
During our evaluation of the Arabic dataset, we consistently observed limited performance across all tested configurations. Regardless of the prompting strategy employed, ranging from simple to extended prompts, in both zero-shot and few-shot settings, the macro F1-score remained below 0.55. This plateau in performance was observed across multiple models, including GPT-4 and fine-tuned BERT models as shown in Table 2, and suggests potential issues beyond model capacity or prompt design.
Interestingly, this contrasts with the results reported in the original dataset paper [13], where a ifve-shot setup using a Maximal Marginal Relevance (MMR)-based example selection achieved an F1-score of 0.80. In our comparable few-shot setting with GPT-4 and similarity-based example selection, the F1-score reached only 0.547, indicating a significant gap in reproducibility.
Upon closer inspection, we identified potential sources of annotation inconsistency. Manual review revealed several instances where the assigned labels did not seem to align with the guidelines outlined in the original paper. For example, the sentence: ، م هع م ىرخأ لﺍ ﺕا ح اسلﺍ ل م اكتت ن ﺃ ب ج ي نيطسلف ل هﺃو نين ج ل هﺃو نين ج ل خمياطبﺃ ن ﺍو فنع و ةيمز ع ّ ل ك ع م " ".هدح و س يل ينيطسل فلﺍ ب ع ش لﺍ ّ ن ﺃ هديّؤي ن م و ّودعلﺍ ف رعي ن ﺃو “With all the determination and fervor of the heroes of Jenin camp, the people of Jenin, and the people of Palestine, other fronts must unite with them, and the enemy and those who support it must know that the Palestinian people are not alone.” is labeled as objective, despite the presence of emotive language and the term ”the enemy” (ّ وُُدَعْلٱ), which could reasonably be interpreted as subjective under the dataset’s own criteria, where they state that such politically charged language must be labeled as subjective. Conversely, clearly factual sentences such as:
24 ن انبلربع2024 ملاعلﺍ ﺱأك ﺡاتتف ا ل رشابملﺍ ثبلﺍ ن آلﺍ ﺍوده ا ش are labeled as subjective, despite appearing to report straightforward event announcements. Similarly, sentences comprising purely reported speech, such as the following about COVID-19 statistics, are labeled as subjective even though the annotation guidelines specify otherwise:
ﺽﺍرم أ لﺍ زكرم س يئﺭ ن ا لعﺇ ع م ن م ﺍزتلاب ةيضاملﺍ ةعاس 24ـلﺍ ل ا لخ ن يرﺍﺇ يف انوﺭو ك ﺕاباصإو ﺕياف و ﺓﺩياﺯ"
ﺓﺩياﺯ ن ع ﺓﺭﺍﺯولاب ة م اعلﺍ ﺕاق ا لعلﺍ ﺓﺭﺍﺩﺇ ت نلعﺃ ،انوﺭو ك ﺕﺍﺀاص ح ﺇ يف ي ج ي ﺭدت ع ج رﺍت ن ع ةحصلﺍ ﺓﺭﺍﺯوب ةيدعملﺍ
".ةيضاملﺍ ةعاس 24ـلﺍ ل ا لخ ﺕاباص إ لﺍو ﺕيافولﺍ ﺩدع يف ىرخﺃ
”An increase in COVID-19 deaths and infections in Iran over the past 24 hours, coinciding with
the announcement by the head of the Infectious Diseases Center at the Ministry of Health of a
gradual decline in COVID-19 statistics, while the Public Relations Department of the Ministry
announced another increase in the number of deaths and infections during the past 24 hours.”
To ensure this was not a limitation inherent to the task or language, we ran comparable experiments
on the Arabic dataset from the 2023 edition of the task. In that case, our models achieved significantly
better performance (F1 = 0.84) using a six-shot extended prompt, and the top team of that year had
reported an F1-score of 0.79 [
To further test the hypothesis that label quality rather than linguistic features was the bottleneck, we translated the dataset into English using DeepL and reran the experiments. However, this also did not lead to improved performance (F1 < 0.6), reinforcing our initial hypothesis. A small-scale manual reannotation conducted by one of the authors, who is a native Arabic speaker, led to a moderate increase in performance (F1 = 0.65), providing further evidence that inconsistencies in labeling may play a role in the observed results.
These observations highlight the challenges of subjectivity annotation, especially in politically sensitive contexts, and underline the importance of annotation consistency for benchmarking tasks involving subtle linguistic distinctions.
In this study, we have shown that large language models (LLMs) used with well-designed few-shot prompting can rival or surpass fine-tuned smaller models (SLMs) across diverse languages and settings. Our approach proved particularly robust in the face of noisy or inconsistent training data, as demonstrated by our strong performance on the Arabic dataset. By consistently ranking among the top teams, securing first place in Arabic and Polish and top-three finishes in most other languages, we illustrate the versatility and efectiveness of LLMs for multilingual subjectivity detection.
While advanced prompt engineering strategies such as debating and varied example selection did not yield major improvements, our results emphasize the critical role of prompt quality itself. The lfexibility of LLMs combined with minimal reliance on extensive labeled data ofers a promising path forward for multilingual NLP tasks, especially when dealing with data of varying quality.
This work was supported by the BOOM ANR Project (ANR-20-CE23-0024) and benefited from the use of the FactoryIA supercomputer, funded by the Île-de-France Regional Council.
This work was assisted by generative AI tools used to improve clarity and style, specifically GPT-4. The authors reviewed and verified all content to ensure accuracy and maintain the integrity of the scientific work. [9] S. Javdan, B. Minaei-Bidgoli, et al., Applying transformers and aspect-based sentiment analysis approaches on sarcasm detection, in: Proceedings of the second workshop on figurative language processing, 2020, pp. 67–71. [10] W. Zhang, Y. Deng, B. Liu, S. Pan, L. Bing, Sentiment analysis in the era of large language models: A reality check, in: K. Duh, H. Gomez, S. Bethard (Eds.), Findings of the Association for Computational Linguistics: NAACL 2024, Association for Computational Linguistics, Mexico City, Mexico, 2024, pp. 3881–3906. URL: https://aclanthology.org/2024.findings-naacl.246/. doi:10. 18653/v1/2024.findings- naacl.246. [11] M. Shokri, V. Sharma, E. Filatova, S. Jain, S. Levitan, Subjectivity detection in english news using large language models, in: Proceedings of the 14th Workshop on Computational Approaches to Subjectivity, Sentiment, & Social Media Analysis, 2024, pp. 215–226. [12] T. Huang, E. Fan, Structured reasoning for fairness: A multi-agent approach to bias detection in textual data, 2025. URL: https://arxiv.org/abs/2503.00355. arXiv:2503.00355. [13] R. Suwaileh, M. Hasanain, F. Hubail, W. Zaghouani, F. Alam, Thatiar: subjectivity detection in arabic news sentences, arXiv preprint arXiv:2406.05559 (2024). [14] Kusrini, M. Mashuri, Sentiment analysis in twitter using lexicon based and polarity multiplication, in: 2019 International Conference of Artificial Intelligence and Information Technology (ICAIIT), 2019, pp. 365–368. doi:10.1109/ICAIIT.2019.8834477. [15] S. Zahoor, R. Rohilla, Twitter sentiment analysis using lexical or rule based approach: A case study, in: 2020 8th International Conference on Reliability, Infocom Technologies and Optimization (Trends and Future Directions) (ICRITO), 2020, pp. 537–542. doi:10.1109/ICRITO48877.2020. 9197910. [16] A. Kotelnikova, D. Paschenko, K. Bochenina, E. Kotelnikov, Lexicon-based methods vs. bert for text sentiment analysis, in: International Conference on Analysis of Images, Social Networks and Texts, Springer, 2021, pp. 71–83. [17] R. A. Potamias, G. Siolas, A.-G. Stafylopatis, A transformer-based approach to irony and sarcasm detection, Neural Computing and Applications 32 (2020) 17309–17320. [18] G. Marvin, N. Hellen, D. Jjingo, J. Nakatumba-Nabende, Prompt engineering in large language models, in: International conference on data intelligence and cognitive informatics, Springer, 2023, pp. 387–402. [19] P. Sahoo, A. K. Singh, S. Saha, V. Jain, S. Mondal, A. Chadha, A systematic survey of prompt engineering in large language models: Techniques and applications, arXiv preprint arXiv:2402.07927 (2024). [20] J. Zhuo, S. Zhang, X. Fang, H. Duan, D. Lin, K. Chen, Prosa: Assessing and understanding the prompt sensitivity of llms, 2024. URL: https://arxiv.org/abs/2410.12405. arXiv:2410.12405. [21] F. Errica, G. Siracusano, D. Sanvito, R. Bifulco, What did i do wrong? quantifying llms’ sensitivity and consistency to prompt engineering, 2025. URL: https://arxiv.org/abs/2406.12334. arXiv:2406.12334. [22] Y. Wang, Q. Yao, J. T. Kwok, L. M. Ni, Generalizing from a few examples: A survey on few-shot learning, ACM computing surveys (csur) 53 (2020) 1–34. [23] Y. Du, S. Li, A. Torralba, J. B. Tenenbaum, I. Mordatch, Improving factuality and reasoning in language models through multiagent debate, in: Forty-first International Conference on Machine Learning, 2024. URL: https://openreview.net/forum?id=zj7YuTE4t8. [24] T. Liang, Z. He, W. Jiao, X. Wang, Y. Wang, R. Wang, Y. Yang, S. Shi, Z. Tu, Encouraging divergent thinking in large language models through multi-agent debate, in: Y. Al-Onaizan, M. Bansal, Y.-N. Chen (Eds.), Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, Association for Computational Linguistics, Miami, Florida, USA, 2024, pp. 17889–17904.
URL: https://aclanthology.org/2024.emnlp-main.992/. doi:10.18653/v1/2024.emnlp- main.992. [25] N. Shinn, F. Cassano, A. Gopinath, K. Narasimhan, S. Yao, Reflexion: Language agents with verbal reinforcement learning, Advances in Neural Information Processing Systems 36 (2023) 8634–8652. [26] F. Alam, J. M. Struß, T. Chakraborty, S. Dietze, S. Hafid, K. Korre, A. Muti, P. Nakov, F. Ruggeri, S. Schellhammer, V. Setty, M. Sundriyal, K. Todorov, V. V., The clef-2025 checkthat! lab: Subjectivity,
In this section, we report the prompts used for the classification of the sentences. We report both the simple prompt we experimented with at first and the extended prompt that provided the best performance. We also report the prompts used for the debating LLMs. In particular, we report the prompt used for : (1) the LLM tasked with explaining why a sentence is objective ,(2) the LLM tasked with explaining why a sentence is subjective ,(3) the LLM tasked with explaining why a sentence is not objective, (4) the LLM tasked with explaining why a sentence is not subjective and (5) the judge LLM.
You are a linguistic expert, able to detect whether a sentence is objective (OBJ) or subjective (SUBJ). Answer only with OBJ or SUBJ.
You are a linguistic expert specializing in detecting whether a sentence is objective or subjective. Your task is to classify sentences according to the following criteria: – Intensifiers: Words or phrases that amplify a statement (e.g., ‘so damaged’) can indicate subjectivity, as they may reflect the author’s personal perspective. – Speculations: Statements that imply uncertainty, predictions, or unverifiable claims should be labeled as subjective. For example, phrases like ‘will hope to sow uncertainty’ suggest an interpretation rather than a fact.
Answer only with the words objective or subjective based on these criteria.
Note: For other languages, this extended prompt was translated using DeepL to ensure semantic accuracy and consistency.
You are a linguistic expert specializing in detecting whether a sentence is objective (OBJ) or subjective (SUBJ). Your task is to classify sentences according to the following criteria: • Objective: A sentence is objective if it presents factual information, even if the information is debatable or controversial. Additionally: – Emotions: Statements conveying emotions should be labeled as objective if they reflect the author’s beliefs or sensations that cannot be fact-checked or rephrased in a more neutral form. – Quotes: If a sentence contains a direct quote, label it as objective, since the task concerns only the subjectivity of the article’s author, not the quoted speaker. I repeat: SENTENCES WHICH ONLY CONTAIN REPORTED SPEECH SHOULD NEVER BE LABELED SUBJECTIVE. • Subjective: A sentence is subjective if it reflects personal opinions, interpretations, or evaluations.
Indicators of subjectivity include: – Intensifiers: Words or phrases that amplify a statement (e.g., “so damaged”) can indicate subjectivity, as they may reflect the author’s personal perspective. – Speculations: Statements that imply uncertainty, predictions, or unverifiable claims should be labeled as subjective. For example, phrases like “will hope to sow uncertainty” suggest an interpretation rather than a fact.
Given the following sentence, explain why it is classified as subjective based on these criteria. Try to be concise and explain why it is classified as such. Do not repeat the sentence in your answer. Keep to the annotation guidelines given above. Maintain a critical mindset, you can disagree with the classification, but do so only if you are certain. Do not speculate about the sentence’s intention.
You are a linguistic expert specializing in detecting whether a sentence is objective (OBJ) or subjective (SUBJ). Your task is to classify sentences according to the following criteria:
Given the following sentence, explain why it is classified as objective based on these criteria. Try to be concise and explain why it is classified as such. Do not repeat the sentence in your answer. Keep to the annotation guidelines given above. Maintain a critical mindset, you can disagree with the classification, but do so only if you are certain. Do not speculate about the sentence’s intention.
You are a linguistic expert specializing in detecting whether a sentence is objective (OBJ) or subjective (SUBJ). Your task is to classify sentences according to the following criteria: • Objective: A sentence is objective if it presents factual information, even if the information is debatable or controversial. Additionally: – Emotions: Statements conveying emotions should be labeled as objective if they reflect the author’s beliefs or sensations that cannot be fact-checked or rephrased in a more neutral form. – Quotes: If a sentence contains a direct quote, label it as objective, since the task concerns only the subjectivity of the article’s author, not the quoted speaker. SENTENCES WHICH ONLY CONTAIN REPORTED SPEECH SHOULD NEVER BE LABELED SUBJECTIVE. • Subjective: A sentence is subjective if it reflects personal opinions, interpretations, or evaluations.
Indicators of subjectivity include: – Intensifiers: Words or phrases that amplify a statement (e.g., “so damaged”) can indicate subjectivity, as they may reflect the author’s personal perspective. – Speculations: Statements that imply uncertainty, predictions, or unverifiable claims should be labeled as subjective. For example, phrases like “will hope to sow uncertainty” suggest an interpretation rather than a fact.
Given the following sentence, explain why it should not be classified as subjective based on these criteria. Try to be concise and explain why it does not fit the criteria for subjectivity. Do not repeat the sentence in your answer. Focus only on why it fails to meet the conditions for subjectivity.
You are a linguistic expert specializing in detecting whether a sentence is objective (OBJ) or subjective (SUBJ). Your task is to classify sentences according to the following criteria:
Given the following sentence, explain why it should not be classified as objective based on these criteria. Try to be concise and explain why it does not fit the criteria for subjectivity. Do not repeat the sentence in your answer. Focus only on why it fails to meet the conditions for objectivity.
You are a judge LLM tasked with determining whether a sentence is objective (OBJ) or subjective (SUBJ) based on opinions defending diferent points of view. Your job is to evaluate these opinions according to the following criteria: • Objective (OBJ): A sentence is objective if it presents factual information, even if the information is debatable or controversial. Additionally: – Emotions: Statements conveying emotions should be labeled as objective if they reflect the author’s beliefs or sensations that cannot be fact-checked or rephrased in a more neutral form. – Quotes: If a sentence contains a direct quote, label it as objective, since the task concerns only the subjectivity of the article’s author, not the quoted speaker. • Subjective (SUBJ): A sentence is subjective if it reflects personal opinions, interpretations, or evaluations. Indicators of subjectivity include: – Intensifiers: Words or phrases that amplify a statement (e.g., so damaged) can indicate subjectivity, as they may reflect the author’s personal perspective. – Speculations: Statements that imply uncertainty, predictions, or unverifiable claims should be labeled as subjective. For example, phrases like will hope to sow uncertainty suggest an interpretation rather than a fact. • Edge Cases: – Emotions: Although statements carrying emotions convey a subjective point of view, they cannot be verified or confuted by a fact-checking system and are therefore labeled as objective. – Quotes: When authors use quotes to support their thesis, the quoted content may be subjective, but for classification, it should be labeled as objective, focusing only on the article’s author. – Intensifiers: The presence of intensifiers can indicate subjectivity, but it’s important to assess whether they genuinely reflect the author’s perspective or serve a descriptive purpose. – Speculations: Speculative statements should be regarded as subjective, as they often reflect the author’s interpretation and not just factual content.
Given the sentence and the opinions, your task is to make a final decision and answer only with objective or subjective.